In November 2023, when DeepSeek released its first-generation model, its KV cache required 389,120 bytes per token—an unexceptional figure at the time. By December 2025, with DeepSeek-V3.2, that footprint plummeted to 48,068 bytes, an eightfold reduction. The experimental V3.2-Exp introduced DeepSeek Sparse Attention (DSA) specifically to make long-context training and inference significantly cheaper. By V4 Preview, DeepSeek titled its announcement deliberately: "Entering the Era of Low-Cost Million-Token Context." The emphasis was not merely on "million-token context," but on "low cost." In April 2026, V4-Flash pushed the cache down to 3,514 bytes. Now, DeepSeek V4.1 Flash slashes the KV cache further to a mere 890 bytes per token.

Source: DeepSeek

Three years in, DeepSeek has delivered exponential efficiency gains. Compared to the previous generation, DeepSeek V4.1 Flash requires only 1/4 of the HBM and 1/8 of the SSD storage.

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is a 552-billion parameter MoE (Mixture-of-Experts) model, serving as the smallest member of DeepSeek’s all-new architecture family. Built on a novel Causal Encoder–Decoder design, it takes an unconventional approach to activated parameters by separating "reading" and "writing" into two distinct tasks with dedicated computational budgets: 8B activated parameters on the input side handling context processing, and 16B activated parameters on the output side handling generation.

This asymmetric allocation is not arbitrary, it reflects the real-world distribution of workloads. When an agent parses a 200-file repository, processes lengthy tool execution logs, or analyzes a stack of screenshots, the reading volume is massive while the writing volume is relatively small. Standard architectures that allocate uniform compute based on the heavier output workload waste compute on every input token. Asymmetric allocation eliminates this inefficiency. Combined with an 890-byte per token cache, the target for this optimization is clear: long-running agents, rather than simple single-turn conversations.

DeepSeek attributes these performance gains to novel pre-training methodologies alongside larger-scale post-training Reinforcement Learning (RL). Notably, the model features native visual understanding—not by retrofitting a vision adapter onto a text LLM, but by engineering multimodal processing directly into the underlying architecture.

All data based on official announcements from DeepSeek on September 10, 2026.

Benchmark Performance

DeepSeek compared V4.1 Flash against its previous in-house checkpoints, V4-Pro (0813) and V4-Flash (0731), as well as industry benchmarks GLM-5.3, Kimi K3, GPT-5.6-Sol, and Claude Opus 5.

Source: DeepSeek

What This Means for Teams Running Agents

Consider an agent executing a repository-wide code refactoring that triggers 60 sequential tool calls. Every execution appends new context to an ever-expanding history payload. Each step produces a few hundred output tokens while retaining tens of thousands of context tokens. Multiplying the KV cache overhead across every token in context reveals where the real operational costs lie.

Slashing HBM usage to 1/4 and SSD footprint to 1/8 is fundamentally different from a marginal reduction in output token pricing. The former doesn't just make individual steps slightly cheaper—it makes executing significantly more steps feasible. For autonomous agents that deliver value only upon task completion, total step capacity is the ultimate bottleneck.

Target Use Cases

• Long-Horizon Software Engineering: Repo-level refactoring, CI failure diagnosis, and complex multi-step backend tasks requiring long tool-call histories. This is highlighted by the model more than doubling V4-Pro's performance on Terminal-Bench 3.0 and 4.0, where the 4x reduction in HBM footprint compounds value across every execution step.
• High-Concurrency Agent Pipelines: Workflows where per-step costs dictate the maximum viable step count. At an 8B input / 16B output activated parameter footprint, V4.1 Flash achieves top scores on Automation-Bench and Agents' Last Exam across the entire DeepSeek lineup.
• Security Engineering: A focused yet valuable entry point. Reaching an 88.1 score on CyberGym (the highest reported in DeepSeek’s benchmark suite), along with gains in SEC-Bench Pro and ExploitGym over V4-Pro. While behind GPT-5.6-Sol, its pragmatic positioning covers initial vulnerability screening, security code audits, and crash analysis, rather than end-to-end exploit synthesis.
• Document and Vision Processing Pipelines: Multimodal workflows where dedicated visual extraction stages can be eliminated entirely.

Run DeepSeek V4.1 Flash via API on Bitdeer AI Model Studio

Deploy and run DeepSeek V4.1 Flash seamlessly on Bitdeer AI Model Studio without managing underlying infrastructure, significantly reducing deployment complexity and accelerating time-to-value. As a Preferred NVIDIA Cloud Partner with ISO/IEC 27001:2022 and SOC 2 Type I & Type II certifications, Bitdeer provides the secure, compliant, and high-performance infrastructure required for production-grade agentic deployments.

How to Get Started

1. Log in to Bitdeer AI Model Studio.
2. Locate DeepSeek V4.1 Flash in the model registry.
3. Generate your API key and start making API calls.

curl -v --location 'https://api-inference.bitdeer.ai/v1/chat/completions' --data '{"model":"deepseek-ai/DeepSeek-V4.1-Flash","messages":[{"role":"system","content":"You are a knowledgeable assistant. Provide concise and clear explanations to scientific questions."},{"role":"user","content":"Can you explain the theory of evolution in simple terms?"}],"max_tokens":200,"top_p":1.0,"temperature":1.0,"frequency_penalty":0.0,"presence_penalty":0.0,"seed":0,"stream":false}' --header 'Authorization: Bearer <API_KEY>'

The Bitdeer AI team had a productive experience at Ai4 2026 in Las Vegas, NV, August 4-6, 2026, connecting with customers, partners, and industry leaders while attending sessions across AI infrastructure, enterprise adoption, AI applications, and emerging technologies. 

An overarching theme of the event was AI innovation is moving beyond just experimentation and model development and towards production scale deployment thus creating an increasing demand for reliable, scalable, and accessible AI infrastructure. 

The conference itself showed a clear progression across three days with the first day focusing on infrastructure constraints limiting AI growth, to the challenges of operationalizing AI and ending with what it takes to deploy AI reliably and responsibly in production. 

Day 1 Highlights

Day 1 focused on infrastructure requirements behind the rapid expansion of AI. In one of the opening keynotes, The AI Reckoning: Chips, Constraints and the Next-Generation of Compute, Pat Gelsinger, Playground Global, emphasized AI innovation is accelerating rapidly but hardware limitations and infrastructure including GPU availability, energy capacity, data center scalability, networking, and hardware innovation are critical limiting factors. The same theme emerged in discussions around AI-driven healthcare and drug discovery. In From Molecule to Market: How AI is Reshaping Drug Discovery session, speakers highlighted how AI is being applied across target identification, molecule design, clinical research and trial optimization helping to accelerate discovery cycles and improve decision making. But on the flip side, the discussion honed in on how access to sufficient compute is becoming a constraint to this AI-driven research. As AI expands into compute-intensive industries including healthcare, scientific research, robotics, and advanced simulation, infrastructure capacity will be influential in how quickly organizations can turn AI innovation into real world outcomes. 

Day 2 Highlights

Day 2 delved into the challenges of organizations figuring out how to operationalize AI. The day kicked off with a keynote, The Architects of Intelligence: A Historic Convergence with speakers that included Geoffrey Hinton, Fei-Fei Li, Andrew Ng, where they explored the broader implications of AI adoption. The conversation surfaced the importance of balancing optimism around AI’s potential with responsible adoption while investing in education and AI literacy to help organizations and communities adapt. 

Following keynotes from Vultr and PayPal highlighted the challenge is no longer AI capability, it is enterprise adoption and operationalization. While frontier AI companies continue pushing model capabilities, enterprises are increasingly focused on how to operationalize AI through scalable infrastructure, inference, AI agents, and secure deployment environments. As model capabilities continue to advance, organizations increasingly need infrastructure, platforms, and operational capabilities required to turn these capabilities into business outcomes. 

Day 3 Highlights

The conference ended with conversations centered on how to make AI reliable enough for production. Successful AI adoption requires a combination of infrastructure, security, governance, and workforce readiness. Across keynotes and sessions from Cisco, Waymo, and Crusoe, the message was clear: building an AI model or prototype is only the beginning. The real challenge is deploying AI reliably at scale. Organizations are more and more focused on: production-ready AI applications and agents, reliable inference and workload performance, operational consistency, continuous evaluation and measurement, scalability and predictable performance and moving from demos to mission-critical workloads. 

Final Takeaway 

Altogether, the conference delivered a clear message: AI is no longer just building more capable models. AI is moving from experimentation to production and the infrastructure supporting it must evolve accordingly, transitioning from providing compute to delivering reliable, scalable, secure, and production-ready AI environments.

For Bitdeer AI, these trends reinforce the importance of providing organizations with flexible infrastructure and cloud capabilities across the AI stack. Bitdeer AI combines AI Datacenter, AI Infrastructure, AI Cloud, and Model Studio / MaaS capabilities to support an organization's full journey from experimentation to production. As enterprises scale AI, access to reliable infrastructure and flexible model-access capabilities allows organizations to accelerate deployment and improve scalability, performance, and operational flexibility without having to build every layer themselves.

People say most gains in frontier AI come from bigger base models. Z.ai just showed that isn't the only path: GLM-5.3 uses the exact same foundation model as its predecessor, GLM-5.2, and still delivers the largest coding and security jumps in the GLM family's history, entirely through scaled post-training. For engineering and security teams running long-horizon coding agents or vulnerability-discovery pipelines, that means a meaningfully more capable model without waiting on a new base model cycle.

Today, Z.ai GLM-5.3, described as "Built to Code. Ready for Cyber Defense," is available on Bitdeer AI Model Studio, bringing one of the strongest open-weight coding and agentic models to our serverless inference platform, built for secure, scalable enterprise AI. You can start calling it through a single API today, without provisioning or managing the underlying GPU infrastructure.

What Is GLM-5.3?

GLM-5.3 is Z.ai's newest release in the GLM-5 family, a Mixture-of-Experts (MoE) foundation model with 744B total parameters and roughly 40B activated per forward pass, using a sparse attention mechanism for efficient long-context inference. Critically, GLM-5.3 reuses the same base weights as GLM-5.2 unchanged. Every reported capability gain comes from an extended post-training pipeline, built on a larger set of task environments, more environment types, and Z.ai's asynchronous agentic reinforcement learning framework.

That approach concentrated gains in two areas: long-horizon coding and cybersecurity. Z.ai reports GLM-5.3 is its most powerful open-weights coding model to date, with the largest jumps on the hardest, most agentic benchmarks, and says the model's vulnerability-discovery ability advanced further than the team expected, with GLM-5.3 beginning to reason across multiple stages of an exploit and form coherent plans for complete exploitation chains.

Key Specifications

Why Post-Training-Only Gains Matter

Enterprises evaluating always-on coding and security agents run into a familiar set of tradeoffs:

New base model vs. faster iteration. Retraining a foundation model is slow and expensive. Z.ai's result shows a mature base can still absorb large capability gains through post-training alone, which means meaningful upgrades can ship faster and more often.

General coding ability vs. long-horizon execution. Many models handle short coding tasks well but degrade over long, multi-step agent runs. GLM-5.3's gains concentrate exactly there, on benchmarks that measure sustained, multi-step execution rather than single-shot code generation.

Open access vs. dual-use risk. GLM-5.3's cybersecurity ability is strong enough that Z.ai delayed the open-weight release to complete what it calls its most extensive risk review to date. That is a meaningful signal for enterprises about the model's real-world capability, and about the seriousness with which Z.ai is treating deployment.

Coding and Agentic Performance

Z.ai reports a 50% improvement over GLM-5.2 on its internal Z.ai Code Bench, and open-source state-of-the-art results on several public agentic benchmarks:

Source: Z.ai

The largest jumps land on the longest-horizon, most agentic tasks, such as Terminal-Bench 3.0, which measures a model's ability to chain shell commands, read output, and recover from errors across many steps. On public leaderboards, GLM-5.3 trails top closed frontier models on some of the hardest coding evaluations, and Z.ai is transparent that its headline 50% figure comes from an internal benchmark designed to reduce contamination risk rather than an independently reproducible public suite.

An Unplanned Cybersecurity Result

Z.ai says it added vulnerability-discovery training data expecting incremental gains in single-bug reasoning. Instead, capability kept compounding as training scaled, and the model began forming coherent, multi-stage exploitation plans rather than single-step findings.

Working with security teams and research groups, Z.ai used GLM-5.3 to find 2,436 vulnerabilities across 269 open-source projects, some in codebases around 40 years old. On benchmarks:

The CyberGym score edges past both Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%), though the margin is close enough that run-to-run variance could account for it. ExploitBench, which requires root-cause reasoning through to a working exploit, more than doubled but still trails Mythos 5's 78.0%, underscoring that the deepest, most ambiguous exploitation reasoning remains a frontier-model strength for now.

Use Cases

Long-horizon coding agents. Repository-scale refactors, CI failure triage, and multi-step backend engineering that require sustained context and error recovery across dozens of tool calls.

Application and product security. Secure code review, crash triage, and white-box vulnerability discovery for teams shipping kernels, browser engines, network stacks, or other security-sensitive software.

DevSecOps pipelines. Automated pull-request review and CI-integrated static and dynamic analysis that flags exploitable patterns before code ships.

Fintech and e-commerce engineering. Long-running agents that maintain complex, multi-service codebases where reliability and instruction-following over many steps matter as much as raw code quality.

MSSPs and security vendors. Structured vulnerability triage, correlation, and reporting at a scale that would otherwise require significant analyst headcount.

Run GLM-5.3 via API on Bitdeer AI Model Studio

You can run GLM-5.3 on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable. With a unified API, our Model Studio lets developers and enterprises start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value, so you can plug a frontier-adjacent coding and security model into your agent stack without standing up new serving infrastructure.

Bitdeer AI is a preferred NVIDIA Cloud Partner, certified to ISO/IEC 27001:2022 and SOC 2 Type I & Type II, providing the secure, compliant, high-performance, enterprise-grade infrastructure that production agentic AI deployments require.

Get Started

1. Log in to Bitdeer AI Model Studio.
2. Locate GLM-5.3 in the model list.
3. Generate an API key and start making API calls.

curl -v --location 'https://api-inference.bitdeer.ai/v1/chat/completions' --data '{"model":"zai-org/GLM-5.3","messages":[{"role":"system","content":"You are a knowledgeable assistant. Provide concise and clear explanations to scientific questions."},{"role":"user","content":"Can you explain the theory of evolution in simple terms?"}],"max_tokens":200,"top_p":1.0,"temperature":1.0,"frequency_penalty":0.0,"presence_penalty":0.0,"seed":42,"stream":false}' --header 'Authorization: Bearer <API_KEY>'

Conclusion

GLM-5.3 is a reminder that frontier progress doesn't always require a new base model, it can come from teaching an existing one to reason further and plan longer.

On August 20, 2026, a mysterious model appeared on OpenRouter.

The model listing contained only a single identifier: stealth/ox-alpha. Its only publicly available information was a 1-million-token context window, native support for text, image, and video input, and unlimited free access during the preview period. 

The result was the largest model launch in OpenRouter's history. Within just a few days, Ox Alpha surged to the top of the platform’s usage leaderboard. Engineers discovered that this mysterious model could complete tasks that were expected to require a frontier model.

Source: OpenRouter

Subsequently, the community began "fingerprinting," and speculation continued for an entire week, ultimately locking onto a "suspect."

On August 26, the mystery was solved. Z.ai announced that Ox Alpha is a brand-new iteration of the GLM series: GLM-5.3-Flash, a Mixture-of-Experts (MoE) model with 320B total parameters and 18B active parameters, and the first natively multimodal member of the GLM-5 series.

Today, GLM-5.3-Flash is officially live on Bitdeer AI Model Studio, our serverless inference platform built for enterprise-grade security and elastic scaling. You can start calling it immediately, without managing underlying infrastructure and without the need to route through anonymous providers.

What is GLM-5.3-Flash?

GLM-5.3-Flash is a Mixture-of-Experts (MoE) model with 320B total parameters and 18B active parameters per token. Z.ai describes it as the first natively multimodal model in the GLM-5 series, built from a newly trained base model rather than post-trained down from the GLM-5.3 flagship, with its architecture and training recipe redesigned around capability-per-unit-of-compute.

Per Z.ai's published results, it outperforms GLM-5.2 across benchmarks and real-world workloads at roughly one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. The weights are published on Hugging Face under an MIT license.

Three architectural changes are worth understanding, because they are the reason the stealth preview held up under real production load:

• Hybrid sparse and linear attention. For the first time in the GLM series, Z.ai combines linear attention, which handles local dependencies efficiently, with sparse attention, which retrieves the globally relevant context. This sharply reduces long-context serving cost while preserving precise long-context behavior.
• IndexPool. At million-token context lengths, retrieval itself becomes the bottleneck. IndexPool compresses groups of indexer key vectors through weighted pooling to hold down latency and memory. Z.ai reports approximately 3× less attention compute and a 4.4× smaller KV cache compared with GLM-5.3.
• Manifold-Constrained Hyper-Connections (mHC). Adopted to improve scaling efficiency, letting the model deliver more capability per unit of activated compute.

Together with a 30T-token multimodal pre-training corpus, these changes are what let a model with only 18B active parameters land within striking distance of frontier coding scores.

Key Specifications

Reported Performance

According to Z.ai, across six coding and agentic benchmarks, GLM-5.3-Flash consistently outperforms GLM-5.2, often by a wide margin: 63.4 vs. 46.2 on DeepSWE v1.1 and 48.8 vs. 26.2 on AutomationBench, while approaching Claude Opus 4.8 overall.

Source: https://z.ai/blog/glm-5.3-flash

Artificial Analysis scores GLM-5.3-Flash at 57 on its Intelligence Index v4.1.1, placing it among the strongest intelligence-per-dollar options currently available in open weights.

Source: Artificial Analysis

Why This Matters for Agent Economics

As enterprises move agentic systems into production, a familiar set of tradeoffs comes into focus. GLM-5.3-Flash addresses them from a specific direction:

Context cost vs. context length. Most models with large advertised windows become uneconomical well before you reach the limit, because attention cost scales against you. The hybrid attention design attacks this directly: roughly 3× less attention compute and a 4.4× smaller KV cache mean a million-token window that is usable in production, not just on a spec sheet.

Multimodal vs. pipeline complexity. Teams building agents that read dashboards, verify UI regressions, or process scanned documents typically stitch together an OCR or vision-to-text stage before the reasoning model. Native image and video input removes that stage, along with the accuracy loss and latency it introduces.

Capability vs. cost. Routing every agent step to a frontier model is expensive when most steps are high-volume and repetitive. A model that approaches frontier coding scores at a fraction of the per-token cost changes which steps you can afford to run at all, and how many iterations an agent can take before finishing.

Openness vs. production readiness. MIT-licensed weights mean the model can be owned, audited, and self-hosted. But a 320B model is a serious deployment project even at 18B active: FP8 weights alone run to roughly 306 GiB before KV cache, on top of tensor parallelism, memory capacity, and serving-stack decisions.

Enterprise Use Cases

Repo-scale coding and engineering agents. Hold an entire codebase in context rather than chunking it, and run long-horizon terminal and refactoring tasks where the agent needs continuity across many turns.

Browser and computer-use agents. Native visual input lets agents interpret what is actually on screen, navigate interfaces, and verify outcomes without a separate vision stage.

Million-token document and log analysis. Contract review, compliance checks, incident forensics, and post-mortem analysis over corpora that would otherwise require a retrieval pipeline and its failure modes.

Visual QA and UI regression checking. Compare screenshots against expected states, describe visual defects, and file structured findings.

Back-office and knowledge work automation. Spreadsheet, deck, and dashboard reasoning where the source material is visual and the task requires structured output.

Run GLM-5.3-Flash via API on Bitdeer AI Model Studio

You can run GLM-5.3-Flash on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable. With a unified, OpenAI-compatible API, Model Studio lets developers and enterprises start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value.

One of the practical advantages of an MIT-licensed open-weight model is portability: the weights are not tied to any single serving environment. Bitdeer AI serves MaaS on the NVIDIA compute infrastructure that we own and operate, in our own data centers. As a preferred NVIDIA Cloud Partner, certified to ISO/IEC 27001:2022 and SOC 2 Type I & Type II, we provide the secure, compliant, high-performance foundation that production agentic deployments require.

Get Started

1. Log in to Bitdeer AI Model Studio.
2. Locate zai-org/GLM-5.3-Flash in the model list.
3. Generate an API key and start making API calls.

curl -v --location 'https://api-inference.bitdeer.ai/v1/chat/completions' --data '{"model":"zai-org/GLM-5.3-Flash","messages":[{"role":"system","content":"You are a knowledgeable assistant. Provide concise and clear explanations to scientific questions."},{"role":"user","content":"Can you explain the theory of evolution in simple terms?"}],"max_tokens":512,"top_p":1.0,"temperature":1.0,"frequency_penalty":0.0,"presence_penalty":0.0,"seed":42,"stream":false}' --header 'Authorization: Bearer <API_KEY>'

Conclusion

Agentic workloads are no longer bottlenecked on intelligence. They are bottlenecked on how much context a team can afford to carry, and on how many pipeline stages sit between the agent and what it needs to see. GLM-5.3-Flash addresses both: 320B total parameters with 18B active, a hybrid attention architecture that makes a 1M-token window economically usable, native image and video input, and MIT-licensed weights you can own.

With GLM-5.3-Flash now available on Bitdeer AI Model Studio, you can start building against it today on secure, enterprise-grade infrastructure, and scale onto dedicated Bitdeer AI GPU capacity when your production volumes call for it.

When building AI capabilities, many teams fall into a common trap: devoting too much attention to the question, "How many GPUs do we actually need?" In enterprise AI infrastructure planning, however, the real challenge is not simply the number of accelerators. It is determining the right resource mix for each workload across development, training, evaluation, and production. Both training and inference rely on accelerated computing, but training prioritizes completion speed and large-scale data movement, whereas inference prioritizes continuous service, response time, concurrency, and long-term cost.

A resource plan should therefore cover CPUs, host memory, GPU memory, storage, networking, scheduling, observability, security, and scaling mechanisms, not merely GPU models and quantities. Identifying the workload first and then defining capacity and deployment methods is usually more reliable than purchasing hardware first and forcing applications to fit it later.

Why AI Training and Inference Need Separate Planning

Training is an intensive, time-bounded compute job. Models repeatedly execute forward and backward passes, and multi-GPU or multi-node training requires frequent synchronization. As a result, training is constrained by GPU memory, accelerator interconnects, storage throughput, and network bandwidth. A shortfall in any layer can leave expensive accelerators waiting for data or communication.

Inference is a long-running production service. It must maintain P50, P95, and P99 latency, throughput, availability, and cost per request under changing traffic. Generative AI, retrieval-augmented generation, and agent workloads may trigger several model calls, retrieval steps, and tool actions for one user request, so capacity cannot be estimated from a single model invocation.

Which Computing Resources Does AI Training Require?

Training capacity should reflect model size, training method, dataset volume, and iteration frequency. Domain-model fine-tuning may only need a small cluster for limited periods, whereas foundation-model, long-context, or multimodal training depends more heavily on large GPU memory, high-speed interconnects, and reliable distributed execution.

Compute, Memory, and Interconnects

Before selecting GPUs, estimate memory for parameters, gradients, optimizer states, activations, and communication buffers. When memory is insufficient, teams can use gradient accumulation, mixed precision, parameter-efficient fine-tuning, or model parallelism, but each technique changes throughput and communication overhead. Multi-GPU plans should also inspect topology so high-end accelerators are not restricted by lower-bandwidth links.

Data Pipelines, Storage, and Scheduling

CPUs handle data loading, preprocessing, logging, and orchestration. Inadequate host memory or slow data loaders reduce GPU utilization even when the accelerators are powerful. Hot datasets can be staged on local NVMe or high-throughput storage, while long-term datasets and model versions can remain in object storage. Checkpoint frequency should balance recovery time against I/O overhead, and quotas, queues, and priorities should prevent teams from competing unpredictably for the same GPU pool.

When enterprises need centralized management of notebooks, training jobs, real-time logs, resource scheduling, framework environments, and team permissions, a unified AI training platform can consolidate these capabilities into one workflow, reducing the burden of deploying and maintaining distributed clusters in-house.

Which Computing Resources Does AI Inference Require?

Inference planning should begin with the service-level objective rather than peak GPU specifications. Offline document processing can tolerate batching, while customer assistants and real-time vision systems require consistently low latency. Define concurrent users, request and output length, timeout rate, availability targets, and peak traffic before choosing instance size.

Context, Batching, and Token Throughput

Language-model memory is used not only for weights but also for the KV cache. KV cache usage typically increases with context length, generation length, and the number of concurrent requests, and is also affected by model architecture and cache precision. The longer the context and the higher the concurrency, the less GPU memory remains for model weights and other requests. Larger batches can improve GPU throughput, but they may also increase queueing time and time to first token. Production systems commonly use continuous batching to organize batches dynamically around request length, concurrency, and latency targets, combined with quantization, caching, and model routing so high-value requests use stronger models while simpler tasks run on smaller models.

Scaling, Availability, and Security

Always-on services should define minimum replicas, warm capacity, scale-out thresholds, queue limits, and fallback behavior. GPU virtual machines are suitable for rapidly creating, releasing, or resizing compute resources. When traffic-based autoscaling is required, they should be combined with container orchestration, load balancing, monitoring metrics, and node-scaling mechanisms. Workloads that are sensitive to performance variance or require dedicated networking or full control of the software stack may use bare metal GPU servers. Containers can lock down model versions, dependencies, and access policies while separating development, testing, and production environments.

Security planning should specify endpoint exposure, authentication, log redaction, data retention, and key management. For sensitive workloads, teams should verify whether data enters shared environments, whether logs contain original prompts, and whether model versions can be audited and rolled back. These controls can be as important as raw inference speed.

How Enterprises Build an AI Resource Planning Process

Step 1: Classify the workload. Separate pretraining, fine-tuning, evaluation, batch inference, real-time inference, RAG, and agent execution. Record model size, input and output length, data sensitivity, and operating schedule.

Step 2: Define measurable targets. Training metrics should include GPU utilization, time to train, recovery time, and cost per run. Inference metrics should include P95/P99 latency, requests per second, cost per million tokens, error rate, cold-start behavior, and scale-out time. Without targets, the organization cannot verify whether capacity is adequate.

Step 3: Match the infrastructure. Elastic clusters fit short-term training peaks, while dedicated bare metal fits stable high-load jobs. Flexible virtual machines suit development, and versioned containers or managed endpoints suit production. API gateways, monitoring, scheduling, databases, and some data-preprocessing tasks can run on general-purpose CPU instances, a virtual private server (VPS), or container services, keeping GPUs focused on accelerated workloads such as model training and inference.

Step 4: Close the capacity and cost loop. Use load tests to determine baseline capacity and safety margins, then allocate training and inference costs by project, department, or application. After launch, review utilization, queue time, traffic growth, model size, instance type, and scaling thresholds regularly.

Common Mistakes in Enterprise AI Resource Planning

Common mistakes include budgeting for training but not long-term inference, comparing GPU specifications without testing storage and networking, sharing one environment across experiments and production, and ignoring model versioning, rollback, and access control. A powerful GPU cannot fix an inefficient data pipeline or replace production governance.

A mature plan helps training finish faster, keeps inference within latency and cost targets, and adapts as models and business demand change. The goal is not to own more GPUs. It is to build a measurable, scalable, and auditable production AI environment.

Which Resource Layers Can an AI Cloud Platform Support?

The value of adopting an AI cloud platform goes beyond easier access to GPUs. It connects compute, storage, networking, model tools, permissions, and deployment workflows. Bitdeer AI Cloud supports a resource path from development to production for teams that need unified management of training, inference, and agent workflows.

During training peaks, teams can expand capacity by job. Once workloads move into inference, teams can use Serverless Model APIs to call text, embedding, image-understanding, and other AI models, reducing the need to build and operate the serving layer in-house. More complex business automation can use an AI Agent Platform to connect models, retrieval systems, and enterprise tools, but the capacity and cost model must still account for the multiple rounds of calls generated by each task.

When an always-on agent stalls or runs up a surprising bill, the instinct is to reach for a bigger, smarter model. But most agent work isn’t frontier reasoning. Always-on agents complete complex, multi-step tasks by gathering context, observing their environment, reasoning over what they know, and acting and every one of those steps calls a language model. Many of those steps are high-volume, repetitive, and domain-specific: exactly the work a smaller, specialized, controllable model can do faster and cheaper than a frontier model.

That is the shift behind systems of models: instead of routing every step to one large model, agents route each step to the right model for the task, frontier capability where it’s needed for orchestration and hard reasoning, and specialized models where the work is high-volume and repetitive.

Today, NVIDIA Nemotron 3.5 Lightning, the fastest open model in its class for always-on agents, is available on Bitdeer AI Model Studio, bringing a fully customizable open model for powering production always-on AI agents to our serverless inference platform, built for secure, scalable enterprise AI. You can deploy it immediately, without managing the underlying infrastructure.

What is NVIDIA Nemotron 3.5 Lightning?

NVIDIA Nemotron 3.5 Lightning is a 30B Mixture-of-Experts (MoE) model with 3B active parameters, distilled from NVIDIA’s frontier Nemotron 3 Ultra model. It is a fully customizable open model designed to power production of always-on AI agents, giving organizations control to own the model, post-train it for specialized tasks, and deploy it wherever their agents run.

Developed with the Nemotron Coalition and built for popular agent harnesses, Nemotron 3.5 Lightning is designed to deliver leading accuracy for coding, tool calling, instruction following, and multi-turn workflows, so agents can complete specialized tasks faster. It is built to slot into a system of models: routers such as NVIDIA NeMo Switchyard can intelligently direct each step of an agent workflow to the best available model in a chosen pool, selecting Nemotron 3.5 Lightning when its accuracy, speed, and deployment flexibility make it the right fit for high-volume specialized work.

The model is built around three core strengths:

• Model control. An open model, trained with open datasets, that enterprises can control, customize, and deploy at the edge, locally, and in the datacenter, while managing model behavior, data handling, and agent workflows.
• Customizable for high accuracy. Trained for popular agent harnesses for strong out-of-the-box accuracy on agentic tasks, and post-trainable for specialized workflows to improve accuracy in specific enterprise domains.
• Fast task completion. Up to 4x higher  throughput and efficient token rollout help always-on sub-agents and personal agents complete more steps and finish specialized tasks faster while lowering inference cost.

Key Specifications

Why “Systems of Models” Now Decides Agent Economics

As enterprises move always-on agents into production, a familiar set of tradeoffs comes into focus:

• Capability vs. cost. Routing every step to a frontier model is expensive when most steps are high-volume and repetitive.
• Capability vs. speed. Larger models add latency to every turn, and always-on agents run many turns.
• Generalist vs. specialist. A single general model rarely matches a model post-trained for a specific enterprise domain.
• Openness vs. production readiness. Open models can be owned and customized, but teams still need tested, enterprise-grade servers to run them at scale.

Nemotron 3.5 Lightning is designed to address these tradeoffs directly:

Own and control the model. As an open model trained with open datasets, Nemotron 3.5 Lightning lets enterprises control model behavior, data handling, and agent workflows and deploy where their agents run, from the edge to the datacenter and cloud.

Specialize for accuracy. Trained for popular agent harnesses, the model delivers strong out-of-the-box accuracy on agentic tasks and can be post-trained for specialized workflows to raise accuracy in specific enterprise domains, so a small model can outperform a general one on the work that matters to you.

Finish faster. High token-generation throughput keeps per-step latency and cost low, letting always-on agents run more steps and complete specialized tasks faster. According to NVIDIA’s preliminary benchmarks, Nemotron 3.5 Lightning delivers up to 4x higher throughput than comparable open models on agentic multi-turn workloads, enabling agents to complete specialized tasks faster.

Enterprise Use Cases

Always-on personal and productivity agents. Power long-running assistants for everyday tasks—managing email, calendars, projects, and bookings, where high-volume, repetitive steps benefit from a fast, controllable model.

Financial services workflows. Run specialized agents for tasks such as extracting data from documents, checking policy rules, monitoring risk signals, and preparing structured summaries.

Cybersecurity operations. Power specialized security agents that enrich alerts, classify incidents, query logs, validate controls, correlate indicators, and prepare structured findings for analysts.

Telecom. Support autonomous networks and proactive customer care triaging network alarms, optimizing configurations, and answering billing questions.

Retail. Enrich product catalogs, resolve inventory and fulfillment exceptions, assist product discovery, and answer order, return, or loyalty questions.

Run Nemotron 3.5 Lightning via API on Bitdeer AI Model Studio

You can run Nemotron 3.5 Lightning on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable. With a unified API, Model Studio lets developers and enterprises start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value, so you can slot a specialized model into a system-of-models agent architecture without standing up new serving stacks.

Bitdeer AI is a preferred NVIDIA Cloud Partner, certified to ISO/IEC 27001:2022 and SOC2 Type I & Type II, providing the secure, compliant, high-performance, enterprise-grade infrastructure that production agentic AI deployments require.

Get Started

1. Log in to Bitdeer AI Model Studio.
2. Locate NVIDIA Nemotron 3.5 Lightning in the model list.

3. Generate an API key and start making API calls.

curl -v --location 'https://api-inference.bitdeer.ai/v1/chat/completions' --data '{"model":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16","messages":[{"role":"system","content":"You are a knowledgeable assistant. Provide concise and clear explanations to scientific questions."},{"role":"user","content":"Can you explain the theory of evolution in simple terms?"}],"max_tokens":4096,"top_p":1.0,"temperature":1.0,"frequency_penalty":0.0,"presence_penalty":0.0,"seed":0,"stream":false}' --header 'Authorization: Bearer <API_KEY>'

Conclusion

Always-on agents don’t need one big model for every step, they need the right model for each step. NVIDIA Nemotron 3.5 Lightning brings a fully customizable open model to that architecture: a 30B MoE with 3B active parameters, distilled from a frontier model, built for popular agent harnesses, and designed for high throughput on specialized, high-volume work. With Day-0 availability on Bitdeer AI Model Studio, you can start routing your agents’ specialized steps to a model you can own, customize, and run at scale today with higher throughput, lower latency, lower inference cost,  and full control over your data and workflows.

AI workflows are undergoing a massive shift: moving from "can we get this model to run?" to "can we reliably deploy it to production?" Today, developers face challenges that extend far beyond model selection and inference code. Enterprise engineering teams must now piece together a sustainable pipeline that seamlessly integrates data processing, model training, inference serving, Agent orchestration, runtime monitoring, and security controls.

This shifting landscape is exactly why AI GPU computing is back in the spotlight. When most AI projects stall between the demo and production phases, the issue is rarely whether the model can be called. The real hurdles are compute stability, environment replication, inference cost management, and service reliability under high concurrency. GPUs have evolved past being mere hardware accelerators for training large models, they now dictate training cycles, inference latency, and maximum concurrent capacity, directly shaping the cost structure of enterprise AI at scale.

Why Do AI Workflows Depend More on GPU Computing Now?

Traditional cloud computing primarily serves general-purpose workloads like websites, databases, CRMs, and API gateways, relying heavily on CPUs for business logic and request scheduling. AI workloads are fundamentally different. Their core computation shifts toward matrix multiplication, tensor operations, attention mechanisms, embedding generation, and multimodal data processing. While GPUs may assist with vector-related tasks in massive vector computing or accelerated retrieval scenarios, not all vector databases and Approximate Nearest Neighbor (ANN) search implementations inherently require GPU acceleration.

This shift is driving the architecture upgrade of AI data centers. These facilities are built around GPUs or specialized AI accelerators, requiring low-latency, high-bandwidth interconnect networks, high-throughput storage, and tight synergy between compute, networking, and storage. In short, bottlenecks in AI workflows are rarely isolated code issues—they are systemic architectural challenges.

Throughout the AI lifecycle, different stages place starkly different demands on GPU compute architecture:

• Model Training: Highly sensitive to GPU VRAM capacity, compute throughput, multi-GPU interconnect bandwidth (such as NVLink), and long-term stability under heavy workloads.
• Model Inference: Prioritizes low latency, high concurrency, throughput per second (Tokens/s), and cost efficiency per request.
• Agent Deployment: Demands "always-on" continuous availability, permission isolation, high availability for tool-calling interfaces, and robust retry mechanisms.

What Role Does AI GPU Computing Play in the Workflow?

GPUs are uniquely suited for AI because the core computations of modern neural networks are inherently parallelizable. Operations like matrix multiplication, convolutions, attention calculations, and gradient updates during training can be broken down into millions of similar, smaller computational tasks executed simultaneously. While CPUs excel at complex control logic, GPUs are built for high-throughput parallel execution.

However, when selecting GPU resources for production environments, enterprises cannot look at the theoretical raw performance of a single card in isolation. Practical production environments require evaluating whether the model fits comfortably within VRAM, whether multi-GPU communication is fast enough, whether storage can continuously feed data without bottlenecks, and whether containers, drivers, CUDA versions, and frameworks are fully compatible. Teams must also consider if the inference service scales dynamically with traffic, and whether task states, logs, access permissions, and billing remain manageable.

Furthermore, high-density GPU systems are also redefining infrastructure requirements. Using GB200 NVL72 discussed in modern GPU cooling design as an example, a single rack can integrate 72 Blackwell GPUs and 36 Grace CPUs. Each GPU consumes more than 1,000 watts at peak load and uses a hybrid cooling architecture led by cold-plate liquid cooling with air cooling as support. In some public specifications, liquid cooling accounts for about 115 kW of an approximately 132 kW rack power profile, while air cooling accounts for about 17 kW. This more accurately reflects the power, cooling, and rack engineering capabilities required by high-density AI GPU systems.

How Does GPU Computing Optimize the AI Lifecycle?

1. Development & Prototype Validation: Flexible GPU Virtual Machines

At this stage, teams rarely need massive GPU clusters. The priorities are fast startup times, environment consistency, and strict cost control. By leveraging GPU Virtual Machine (VM) instances, developers can quickly spin up environments pre-configured with PyTorch, TensorFlow, CUDA, JupyterLab, or common inference frameworks. This allows teams to rapidly test data pipelines, RAG architectures, model APIs, and Agent logic while eliminating local environment discrepancies, helping enterprises validate AI concepts faster.

2. Model Training & Fine-Tuning: Scalable Distributed Tasks

As models grow larger and datasets expand, the demands on VRAM capacity, compute throughput, network interconnects, and storage I/O scale exponentially. The value of dedicated distributed training task management lies in transforming experimental code into a structured engineering workflow. It automates job submission, resource allocation, health monitoring, log tracking, access control, and cost governance minimizing hidden expenses caused by training failures, environment drift, and idle compute resources.

3. Inference & Model Serving: Highly Elastic Serverless Endpoints

Once a model enters the serving phase, the objective shifts toward achieving low latency, high concurrency, stable APIs, hot-swappable models, and predictable per-call costs. Utilizing Model Serverless Endpoints allows developers to consume text generation, image generation, and visual comprehension models directly via APIs. Compared to building and maintaining an inference cluster from scratch, this approach is ideal for rapid feature validation and scales seamlessly alongside business growth.

4. Agent Deployment & Orchestration: High-Availability Cloud Runtimes

During Agent deployment, the infrastructure needs to support more than model calls. It also needs tool connections, access control, log tracing, and background execution. It is worth noting that the base deployment of Agent orchestration or cloud runtime environments such as OpenClaw usually does not require a GPU. A stable CPU VM can handle always-on operation, tool integration, and background tasks. The value of GPUs is more often reflected in the underlying model inference, embedding generation, or multimodal workloads. Relying on a local laptop to run Agents over the long term can create risks such as shutdown interruptions, complex environment configuration, and API key exposure. With OpenClaw cloud deployment, teams can place cloud instances, model configuration, communication tool integration, and background execution in a more stable environment.

What Architecture Should Different AI Workloads Choose?

Different stages of an AI workflow should not use the same resource strategy. Development validation is well suited for GPU virtual machines; high-performance training or dedicated inference is better suited for bare metal; reproducible deployment is better suited for container services; rapid inference integration is better suited for an AI model library and Serverless Models; large-scale training is better suited for distributed training jobs; and business automation is better suited for an AI Agent platform.

This layered selection is more important than simply pursuing the highest-spec GPU. What enterprises truly need is an upgrade path that can move step by step from prototype to training, inference, and Agent automation. In the early stage, teams can validate model and business assumptions with lighter resources. After entering production, they can gradually introduce dedicated GPUs, containerized deployment, distributed training, and Agent workflows.

How Does Bitdeer AI Cloud Support the End-to-End AI Workflow?

Bitdeer AI Cloud is structured around the complete AI lifecycle. It supports development through deployment and provides two major capability groups: GPU Cloud Services and AI Studio & AI Solutions.

At the infrastructure layer, Bitdeer AI Cloud provides virtual machines, bare metal, and container services. At the training layer, it provides distributed training job capabilities to help teams manage training jobs, monitor status, and review logs. At the inference layer, Model Studio and Serverless Models allow developers to call models through APIs. At the application layer, the AI Agent platform further connects model capabilities to enterprise tools and business processes.

For teams tackling cutting-edge workloads, Bitdeer AI Cloud provides high-performance infrastructure featuring NVIDIA GB200 NVL72 clusters. These systems are purpose-built to handle massive distributed training, real-time inference, multi-agent AI ecosystems, high-performance computing (HPC) simulations, and large-scale data analytics. For engineering teams requiring ultra-dense GPU deployments, ultra-low latency networking, and advanced liquid-cooling infrastructure, these clusters deliver the stability, power, and efficiency needed for production-grade AI applications.

Conclusion

The meaning of AI GPU computing has expanded from accelerating model training to supporting the full AI lifecycle. From development validation to training and fine-tuning, from inference services to Agent deployment, and then to monitoring, scaling, and cost optimization, GPU cloud platforms are increasingly determining whether AI projects can truly enter production.

The next stage of AI competition will not be only a competition in model capability. It will also be a competition in infrastructure efficiency. Enterprises need more than larger GPUs. They need a workflow architecture that allows compute, models, deployment, and Agents to scale together.

Bitdeer AI Cloud serves as the unified gateway for this architectural transition. It enables your teams to start lean with prototype development and smoothly transition into sustainable, production-grade AI workflows, turning isolated experiments into resilient, revenue-generating business systems.

When an AI agent gives a wrong answer, it’s easy to blame the model's reasoning. But in production agentic systems, the failure often happens one step earlier—with retrieval. If the retrieval layer misses the right passage, code file, policy, or customer-specific record, the downstream agent starts from the wrong context and no amount of reasoning power can recover from a bad starting point.

And agents retrieve constantly. They decompose tasks into multiple queries, rewrite queries until the right data surfaces, search memory, and inspect code often dozens of times within a single task. Every weak retrieval adds turns, tokens, latency, and hallucination risk. As enterprises move from simple semantic search to retrieval-driven AI systems, retrieval quality has become the control point for accuracy, cost, and trust.

Today, NVIDIA Nemotron 3 Embed 8B and NVIDIA Nemotron 3 Embed 1B are available on Bitdeer AI Model Studio, bringing production-ready open embedding models for RAG, enterprise search, code retrieval, and agentic workflows. You can immediately deploy the models through Bitdeer AI Model Studio’s serverless inference platform, built for secure, scalable enterprise AI.

What is NVIDIA Nemotron 3 Embed?

Nemotron 3 Embed is a collection of open embedding models designed for enterprise and ISV teams building retrieval systems for enterprise search and emerging agentic workflows. The launch pairs two models on a single, practical accuracy-efficiency curve:

• Nemotron 3 Embed 8B — a frontier-quality embedding model designed to establish the accuracy ceiling across major retrieval benchmarks. The model topped the RTEB leaderboard overall across both open and closed embedding models.
• Nemotron 3 Embed 1B — an efficient embedding model designed to retain more than 95% of the 8B model's accuracy through pruning, distillation, and quantization-aware training (QAT), built for high-volume production workloads that need strong accuracy with substantially better throughput and serving efficiency.

The design philosophy: use the 8B model where maximum retrieval quality matters, and the 1B model where high-volume workloads demand efficiency. Together, they let teams improve retrieval quality without impractical compute tradeoffs.

Key Specifications

Why Retrieval Quality Now Decides Agent Quality

Most enterprise retrieval stacks face a set of hard tradeoffs:

• Accuracy vs. cost. Better retrieval usually means paying more per query.
• Accuracy vs. model size. Larger embedding models can improve retrieval quality, but they are harder to serve at scale.
• Latency vs. recall. Production systems need fast query response, but high-quality retrieval often means searching large volumes of content.
• Openness vs. production readiness. Open models can be customized, but teams still need tested deployment paths, quantization support, and enterprise-grade serving.

This tradeoff is most visible in agentic systems. Multi-turn agents retrieve repeatedly for planning, memory, code, and tool-use context. Weak retrieval increases turn count, token usage, hallucination risk, and user frustration. Strong retrieval reduces irrelevant context and keeps agents grounded. For retrieval to become a default agent capability, embedding models need to feel almost as easy to use as file search: fast, inexpensive, accurate, and available without major compute tradeoffs.

Nemotron 3 Embed addresses these tradeoffs head-on:

Frontier accuracy for retrieval agents. Nemotron 3 Embed models deliver frontier retrieval accuracy across enterprise search, RAG, code retrieval, and agentic retrieval workflows. Nemotron 3 Embed 8B tops the RTEB leaderboard over other open and closed embedding models, and is designed to achieve the highest agentic retrieval accuracy with the fewest tokens used among open embedding models. 

Efficiency without impractical compute tradeoffs. Nemotron 3 Embed 1B retains more than 95% of the 8B model's accuracy through pruning, distillation, and quantization-aware training, and NVFP4 support on NVIDIA Blackwell doubles throughput. This makes semantic search fast and inexpensive enough for agents to retrieve constantly, almost like file search. On NVIDIA Blackwell GPUs, the NVFP4 version delivers 2x higher throughput while retaining 99% of BF16 accuracy.

Open and transparent, from weights to recipes. Nemotron 3 Embed ships with open model weights, datasets, training recipes, optimization techniques, and fine-tuning guidance. Teams can inspect how the model was built, fine-tune it for domain-specific retrieval, and deploy it with full control over their data and infrastructure: no black box, no lock-in. 

Enterprise Use Cases

Agentic retrieval. Give agents a stronger retrieval layer for multi-turn planning, tool use, memory, and repeated lookup loops. Nemotron 3 Embed supports query decomposition where an agent breaks one user request into multiple retrieval queries and query rewriting, where an agent reformulates a query multiple times until the right data is retrieved.

RAG and enterprise search. Ground copilots and RAG pipelines in the right enterprise context documents, knowledge bases, and policies with frontier retrieval accuracy that reduces irrelevant context and wasted tokens downstream.

Code retrieval. Retrieve relevant source files, functions, and implementation examples for developer copilots and software engineering agents. With text and code embedding in one model, engineering teams can power code search and coding-agent context from a single retrieval layer.

Run Nemotron 3 Embed via API on Bitdeer AI Model Studio

You can run Nemotron 3 Embed on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable. With a unified API, Model Studio allows developers and enterprises to start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value.

Bitdeer AI is a preferred NVIDIA Cloud Partner, certified to ISO/IEC 27001:2022 and SOC2 Type I & Type II, providing the secure, compliant, high-performance, and enterprise-grade infrastructure that production retrieval and agentic AI deployments require. Run your embedding workloads at the precision and scale your business requires, on Bitdeer AI's purpose-built GPU fleet.

Get Started

1. Log in to Bitdeer AI Model Studio
2. Locate NVIDIA Nemotron 3 Embed 8B or NVIDIA Nemotron 3 Embed 1B in the model list

3. Generate an API key and start making embedding API calls

NVIDIA Nemotron 3 Embed 1B:

curl -v --location 'https://api-inference.bitdeer.ai/v1/embeddings' --data '{"model":"nvidia/Nemotron-3-Embed-1B-BF16","input":"The cat danced gracefully under the moonlight, its shadow twirling like a silent partner."}' --header 'Authorization: Bearer <API_KEY>'

NVIDIA Nemotron 3 Embed 8B

curl -v --location 'https://api-inference.bitdeer.ai/v1/embeddings' --data '{"model":"nvidia/Nemotron-3-Embed-8B-BF16","input":"The cat danced gracefully under the moonlight, its shadow twirling like a silent partner."}' --header 'Authorization: Bearer <API_KEY>'

Conclusion

Retrieval is becoming the control point of enterprise AI,  the layer that decides whether agents stay grounded or go off course, and whether every reasoning cycle is spent on the right context or wasted on the wrong one. NVIDIA Nemotron 3 Embed brings frontier retrieval accuracy to practical enterprise deployment: an 8B model that sets the accuracy ceiling, a 1B model that carries that quality into high-volume production, and full openness from weights to recipes. With Day-0 availability on Bitdeer AI Model Studio, you can start building higher-quality RAG, search, and agentic retrieval systems today with better context, fewer wasted tokens, and more grounded AI systems.

Authors: Eric Si, Evelyn Xiong

The Two Kinds of AI Nobody Talks About Together

AI has already made several decisions on your behalf today, most of them without you noticing. The sequence your feed showed you. Whether a transaction cleared. None of those involved a chat interface. They operated as embedded decision systems, integrated into existing infrastructure, acting in real time. 

Then there is the AI that has become visible: a text box waiting for a question, a tool you interact with deliberately. Both categories are real and both create value. But they represent fundamentally different levels of complexity, risk, and organizational readiness to build. 

Most public conversation focuses on the second type. Most durable enterprise value is being built with the first. 

Five Technologies Worth Understanding Right Now

The AI landscape is expanding faster than most organizations can track. New model architectures, deployment patterns, and tooling categories emerge constantly, and the full list of technologies reshaping business operations goes well beyond what any single piece can cover. What follows are five capabilities I consider foundational to serious enterprise AI deployments right now. They are not the only ones worth understanding, not by a wide margin, but they are the ones I come back to most when thinking about what it actually takes to move from prototype to production. Think of them as sequential building blocks: each one extends the reach of the last.

Generative AI. This is where most enterprise deployments begin. It produces text, code, summaries, and structured content on demand. In practice it almost always means a working draft that a person reviews before anything goes anywhere. The human review step is not a limitation to be removed. It is the right design for high-stakes output, and organizations that treat it as temporary tend to learn why it was there. 

Retrieval-Augmented Generation (RAG). General-purpose models are trained on general knowledge. They have no awareness of your organization's policies, products, or procedures. RAG solves this by retrieving relevant documents at inference time rather than relying on what was baked into training. The difference between a system that fabricates a plausible answer and one that cites the correct policy is not about which model you licensed. It is a data architecture decision. 

AI Agents. An agent shifts AI from answering to acting. Given a goal and available tools, it plans and executes through to completion. A system that describes available flights is not an agent. A system that searches availability, checks your calendar, completes the booking, and sends a confirmation is an agent. The difference is architectural, not cosmetic.

Multimodal AI. When AI can work across images, scanned documents, audio, and video, entirely new workflows become automatable. An invoice arrives as a photograph. The system reads it, extracts line items, matches them to budget codes, and routes to the right approver with no manual data entry. That removed step is where operational leverage compounds.

AI Governance. Governance is undervalued early and recognized as essential after the first serious incident. The controls, audit mechanisms, and accountability structures governing an AI system are not friction on deployment. They are what makes it possible to grant AI more authority over time without increasing risk. Organizations that treat governance as optional tend to find out why it is not. Each successive capability gives the system more autonomy to act. Governance is what earns the right to grant it.

What Is Actually Inside an AI System

When an AI system fails, the instinct is to blame the model: buy a better one or switch vendors. That is the wrong diagnosis most of the time, and an expensive one to act on. A production AI system has at least five distinct layers, and failure can come from any of them.

Interface layer: What users see. It is designed to appear simple, which means everything complex is hidden behind it. Apparent simplicity is not evidence of underlying simplicity. 

Orchestration layer: The decision engine behind every interaction. It determines whether a message needs a search, whether prior context is relevant, or whether a tool should be called. Invisible in normal use, but it governs how the entire system behaves. 

Knowledge layer: Where the system draws its understanding of your organization. A sophisticated model working from outdated or poorly structured data will consistently produce poor outputs. This is the most common source of failure in enterprise AI, and reliably the last place teams look. 

Tools layer: Without tools, AI can only advise. With tools, it sends messages, updates records, and triggers downstream systems. Agentic capability lives here, as do the consequences when something misfires. 

Human oversight layer: Where accountability is defined and enforced. For routine tasks, AI drafts and a human approves. For consequential decisions involving medical, financial, or legal exposure, a person must sign off before any action is taken. A well-designed system treats this layer as mandatory. When something goes wrong, the right question is not which model to replace. It is which layer broke, and why.

The Progression from Assistant to Agent

AI deployments vary widely in capability and in the consequences of failure. Four stages help frame the difference, because the risk profile escalates meaningfully at each step.

Stage one: AI assistant. Responds when asked. No autonomy, nothing happens without a prompt. This is where most organizations start, and where many stay.

Stage two: AI workflow. Activates when an event occurs, not when a user asks. An invoice arrives; the system extracts data, classifies the expense, checks it against budget, and routes it for approval. Humans are involved at decision points, not every step.

Stage three: AI agent. Receives a goal and determines its own path. Identify strong candidates, review backgrounds, draft outreach, flag anything worth a second look. The objective is given once. The agent works out the steps.

Stage four: multi-agent system. Multiple specialized agents working in coordination: research, compliance, handoffs, oversight, planning. This architecture scales in ways a single-agent system cannot, and it is already in production in certain industries.

The distance between stages carries real consequences. A flawed summary costs minutes to fix. An agent that sends erroneous communications to thousands of customers, misroutes a payment, or cancels a booking without authorization is not making an error. It is creating a business incident with financial, legal, and reputational impact. 

Before deploying at stage three or four, any organization should answer five questions honestly: Is the system's reasoning reliable enough for this specific task? What sensitive data will it access, and does it need all of it? Which systems can it reach, and should it reach them? Which decisions require human approval before action? And if something goes wrong, is the audit trail sufficient to reconstruct what happened? These are not compliance formalities. They are the engineering questions that separate a production-grade system from a well-dressed prototype.

The Skills That Compound Over Time

When AI can produce competent output at near-zero marginal cost, routine output loses scarcity value. Judgment gains it: domain expertise, contextual reasoning, the ability to take responsibility for a decision. Those are precisely what AI does not supply. 

The professionals positioned well in AI-augmented organizations are not those who have adopted the most tools. They are those who understand their domain well enough to recognize when AI output is wrong. Eight skills tend to compound in this environment:

Problem decomposition. Taking an ambiguous situation and breaking it into steps a system can reliably follow. This matters at every level of AI system design, from how instructions are written to how workflows are sequenced.

Workflow design. Knowing which steps to automate, which to augment, and which to keep under direct human control. Where humans stay in the loop is a design decision with real consequences, not a default to leave unexamined.

Critical evaluation. Treating AI output as a starting point, not a conclusion. Every serious AI platform includes an accuracy caveat for good reason. Professionals who internalize it hold an edge over those who do not.

Data literacy. Understanding where information comes from, how it was structured, and when it becomes unreliable. The quality ceiling of any AI system is set by its data, regardless of how capable the model is.

Risk awareness. Thinking through failure modes before they occur. Practitioners with operational AI experience develop intuition for where gaps tend to appear, and that intuition has real market value.

Cross-functional communication. Making complex systems legible to people who did not build them. Every AI system eventually needs to be governed or operated by someone outside the technical team.

Governance thinking. Asking not only what a system can do, but who is accountable when it errs, what data it should access, and whether access controls are actually calibrated to the real risk. 

Continuous learning. The tools and models in use today will look different within two years. Frameworks for understanding systems carry forward. Fluency with any particular tool's interface does not.

The Scarcest Resource in AI Deployment

What remains scarce is not only infrastructure, such as GPU compute, but also the judgment to design AI systems that hold up in production: knowing where to place intelligence within a workflow, where to keep humans accountable, what failure looks like before it happens, and how to build something that performs inside a real organization rather than only in a controlled demonstration. 

To bridge this gap, Bitdeer AI provides tiered resource services for complex AI workflows. The bottom layer is compute, where GPU clusters and data centers form the foundation for all upper-layer functionalities to operate. The middle layer is the AI Cloud, enabling enterprises to host and scale models without the need to manage their own hardware. The top layer is the model application layer, where AI applications connect to live systems and execute multi-step processes.

The shift from AI as a standalone tool to AI as an integrated component of governed, production-grade systems is underway across every sector. Organizations that build strong system design practices early will hold an advantage that is difficult to replicate. 

AI development is no longer limited by model performance alone. For developers and technical leaders, the real production challenge is whether an idea can move from a notebook or small-scale experiment into a secure, repeatable, and scalable GPU cloud environment. Large language model (LLM) training, fine-tuning, Retrieval-Augmented Generation (RAG), batch inference, and low-latency model serving each place distinct demands on VRAM, storage throughput, networking fabric, runtime consistency, access control, and cost visibility. When these infrastructure elements are an afterthought, teams encounter volatile GPU utilization, sluggish deployment cycles, and spiraling operational costs as workloads scale.

At Bitdeer AI, we see the bottleneck as infrastructure orchestration: acquiring suitable GPUs infrastructure, maintain high-speed cluster interconnectivity, continuously feed accelerators via high-throughput storage, manage deterministic runtime environments, and deliver predictable compute for workloads that may scale from a single prototype to distributed production. Our mission is to simplify the infrastructure layer, allowing teams to spend less time managing compute environments and more time building core AI applications.

Why AI GPU Infrastructure Became Harder to Operate

Taking AI workloads to production requires far more than raw compute power. It requires a coordinated system of GPU capacity, CPU orchestration, memory bandwidth, storage throughput, network performance, scheduling discipline, security, and operational control. The gap between a local proof of concept (PoC) and a production AI cloud environment often becomes visible only when teams begin to scale.

GPU Cloud Availability vs. Workload Matching

The primary challenge is not just sourcing GPUs, but precisely matching accelerator configurations to specific workloads. Training, fine-tuning, batch inference, and real-time serving have wildly different requirements for VRAM capacity, throughput, latency, and utilization. Choosing infrastructure before defining workload patterns often forces teams to pay a premium for idle capacity or face unstable delivery due to under-provisioned compute.

Cluster Networking and Data Flow

Modern AI rarely runs as an isolated single-card task. Distributed training relies heavily on fast parameter exchange (via efficient cluster interconnects), fast checkpointing, and storage systems capable of continuously feeding the GPUs. Conversely, inference serving requires orchestrating request routing, dynamic batching, and model serving without letting the network become a bottleneck. The core challenge is not just owning GPUs, but making multiple components operate seamlessly as a unified compute system.

Power Density, Cooling, and Reliability

AI infrastructure also changes the power equation. High-density GPU clusters cannot be planned as if power delivery, cooling, and AI data center design are background concerns.In our infrastructure roadmap, we plan to aggressively expand our IT load dedicated to AI computing, backed by a broader global power infrastructure strategy. For developers, the operational takeaway is clear: reliable AI cloud capacity depends as much on power availability and thermal readiness as it does on the raw accelerator count.accelerator count.

What “Simplifying AI Infrastructure” Actually Means

Simplifying infrastructure does not mean hiding critical technical decisions. Rather, it means stripping away repetitive setup overhead while preserving the controls that truly matter in production: environment consistency, resource elasticity, workload observability, and the ability to transition from development to deployment without rebuilding the tech stack every single time.

Accelerated Environment Provisioning

AI teams should not have to manually stitch together drivers, frameworks, access rules, and runtime dependencies every time they test a workload. A mature GPU cloud workflow bypasses this friction via reproducible environments, making experiments highly portable and production readiness predictable.

Elastic GPU Cloud Capacity

AI compute demands are inherently non-linear. A prototype might require minimal compute, a training job might require a massive cluster for a short burst, and inference serving must scale dynamically with traffic. The true value of cloud GPU infrastructure lies in enabling teams to right-size resources around workload phases, rather than forcing projects to conform to fixed hardware constraints.

Workflow-Level Orchestration

As workloads scale, orchestration becomes the hallmark of infrastructure quality. Containerized tasks, model artifacts, data placement, observability, and deployment pipelines must flow smoothly under reproducible operational controls. Efficient orchestration minimizes friction between research, engineering, and MLOps, the exact gap where AI delivery typically slows down.

How We Transform Infrastructure Into an AI Cloud Platform

At Bitdeer AI, we view GPU cloud infrastructure as an integrated production system, not a fragmented catalog of isolated compute instances. Bitdeer AI Cloud delivers turnkey GPU cloud infrastructure for both training and inference, helping teams align compute choices with the technical DNA of their AI projects. Our goal is to eliminate setup friction, clarify deployment pathways, and bring compute design into lockstep with AI delivery.

GPU Cloud for Training

AI training is hyper-sensitive to VRAM capacity, data throughput, and distributed coordination. We help teams think across the entire training lifecycle: how data reaches the GPUs, how jobs scale across nodes, how environments maintain absolute consistency, and how engineers smoothly transition from experimentation to reproducible production pipelines.

Inference, Model Serving, and Model Studio

Inference introduces an entirely different set of infrastructure challenges, shifting the focus to responsiveness, concurrency, endpoint stability, and the ability to rapidly evaluate models before application integration. Model Studio supports this shift by simplifying model access and offering API-driven inference workflows, allowing teams to evaluate and integrate models with minimal operational overhead.

Security, Cost Predictability, and Operational Fit

For enterprise AI, technical performance alone is insufficient. Teams require workload isolation, reproducible access controls, cost visibility, clear resource planning, and cloud consumption models that align with how AI systems are actually built. These elements become critical as models transition from internal pilots to customer-facing or business-critical services.

Who Benefits the Most?

This infrastructure paradigm delivers maximum value to organizations ready to move beyond isolated experimentation:

Teams Transitioning from Research to Production

Engineering teams that need a clear, managed path from notebooks and early-stage experiments to reproducible training jobs and controlled deployment environments. Their core question is not just GPU access, but maintaining continuity across experimentation, validation, and release.

AI Product Teams Operating High-Scale Inference

Product teams focused on sustained service performance, cost visibility, and the flexibility to scale model capacity without re-architecting the entire system. An AI cloud platform becomes invaluable when it makes inference endpoints easier to evaluate, run, and scale.

Global Teams Building Agentic and Multimodal Workflows

AI agents, multimodal pipelines, and complex enterprise workflows place immense pressure on AI infrastructure, compute availability, and system orchestration. Because these workloads frequently combine inference, retrieval, tool calling, and model serving, the underlying GPU infrastructure, orchestration layer, and API design become core components of the application architecture itself.

The Takeaway: Production-Grade AI Demands Systems, Not Just GPUs

The central question is no longer whether AI teams need accelerators. They do. The more important question is whether the surrounding infrastructure makes those accelerators easier to use, easier to scale, and easier to connect to real deployment requirements.

What AI Teams Should Evaluate

When evaluating an AI cloud platform, teams should look beyond headline GPU availability. Factors like accelerator fit, memory bandwidth, interconnect strategy, storage throughput, orchestration models, inference workflows, security posture, operational visibility, and power-aware infrastructure planning will ultimately dictate how fast an AI system achieves production readiness.

Where Bitdeer AI Fits

Our focus at Bitdeer AI is to make this complex stack accessible through Bitdeer AI Cloud, Model Studio, and continued engagement with the NVIDIA ecosystem. We are not trying to reduce AI infrastructure to a buzzword.We are here to eliminate avoidable setup friction, empowering developers and enterprises to transition from experimentation to production with predictable performance, rigorous deployment discipline, and a clear path to scale.

Over the past year, artificial intelligence has rapidly transitioned from experimentation to real-world deployment. What was once considered an exploratory capability is now being embedded into core business processes across industries. Yet, despite the acceleration in model capabilities, a more fundamental challenge is emerging.

The question is no longer whether AI works.

It is whether organizations can make it work at scale.

In a recent roundtable hosted by Bitdeer AI, discussions with Singapore AI-native companies, business leaders , and technology experts revealed a consistent pattern: while AI capabilities are advancing quickly, deployment maturity is lagging behind. The gap between potential and execution is now the defining constraint.

Where AI Is Delivering Real Value

AI adoption today is increasingly driven by measurable improvements in quality and efficiency, rather than experimentation or curiosity. The strongest traction can be observed in domains where outputs are structured, repeatable, and directly tied to performance metrics.

In areas such as software development, customer support, and sales operations, AI has already demonstrated tangible productivity gains. In one example discussed, AI-enabled sales workflows significantly outperformed traditional approaches, with individual contributors achieving markedly higher output (“20X more efficient”) . While such figures should be interpreted cautiously, they illustrate a broader trend: AI is amplifying individual productivity in ways that were previously unattainable.

This pattern is even more pronounced in data-intensive environments. In auditing workflows, traditional human processes are inherently constrained by time and cognitive bandwidth, often capturing only a fraction of potential issues . By contrast, AI systems can process entire datasets and identify patterns across documents at scale. The result is not merely faster execution, but fundamentally deeper coverage.

However, adoption is not uniform. In creative domains such as image generation, human oversight remains critical, as off-the-shelf model outputs often fall short of production-grade expectations. This uneven distribution of value highlights an important principle: AI adoption does not happen everywhere at once—it concentrates first in areas where value is both immediate and measurable.

A New Model of Execution

AI is not only changing what companies build—it is changing how they build.

Traditional product development cycles, often measured in months, are being replaced by rapid iteration loops measured in weeks. Teams are prioritizing speed over perfection, launching “good enough” solutions early and refining them based on real-world feedback.

This shift is enabled by AI’s ability to lower the cost of experimentation. Building and testing new workflows is no longer prohibitively expensive, allowing organizations to iterate more aggressively and adapt more quickly. The implication is significant. Competitive advantage is no longer determined solely by planning or strategy, but by execution speed and adaptability.

Beyond operational efficiency, AI is increasingly being positioned as a driver of business value and market positioning.

One particularly notable trend discussed was the role of AI in pre-IPO transformation. Some companies are actively integrating AI capabilities prior to listing, using it to strengthen their narrative and enhance perceived growth potential (“AI transformation… higher valuation”) . In this context, AI becomes more than a tool—it becomes part of the company’s identity.

This signals a broader evolution in how AI is perceived:

It is no longer just an operational upgrade. It is a strategic asset.

The Organizational Constraint

Despite clear success in specific use cases, scaling AI across organizations remains challenging. At the organizational level, the barrier is rarely technical — it is structural. But even where alignment is achieved, a second and often more persistent challenge emerges as systems move into production.

AI initiatives are often driven from the top down, with leadership recognizing the strategic importance of adoption. However, this top-down momentum is frequently misaligned with incentives at other levels of the organization. Middle management may resist changes that reduce their operational control, while employees may perceive AI as a threat rather than an enabler.

This misalignment creates a familiar pattern: organizations successfully launch pilot projects but struggle to extend them into production. AI remains siloed, rather than becoming an integrated capability. Even in organizations where alignment exists, deployment introduces a different set of challenges.

From Models to Systems: Cost as the Dominant Constraint

Another key insight from the discussion is that AI deployment is not simply about selecting the right model. It is an ongoing engineering discipline that requires iteration, evaluation, and system design.

Teams described a progression that reflects increasing maturity: initial experimentation with prompting evolves into structured evaluation frameworks, followed by investment in context engineering, workflow orchestration, and ultimately model optimization. This process is rarely linear. Instead, it is characterized by continuous experimentation and refinement .

In this context, the role of engineering shifts. The challenge is no longer to make AI work in isolation, but to make it reliable, consistent, and scalable within production environments. Context management, memory design, and workflow integration become as important as the model itself. Among all these challenges, cost emerges as the most significant constraint.

Token consumption scales with usage, and for AI-native companies, costs can escalate rapidly. What begins as a manageable expense during experimentation can quickly become unsustainable at scale.

To address this, organizations are adopting increasingly sophisticated strategies. Rather than relying on a single model or provider, they are building hybrid architectures that balance performance and cost. Complex reasoning tasks are routed to high-performance models, while repetitive or well-defined workflows are handled by fine-tuned local models. In parallel, some teams are exploring edge deployment to reduce dependency on centralized infrastructure.

At the same time, optimization efforts extend beyond model selection. Teams are actively managing context length, compressing inputs, and building abstraction layers to avoid vendor lock-in. These approaches reflect a broader shift: AI deployment is increasingly a problem of cost-efficient orchestration.

What Enterprises Need Next

As AI systems become more complex, enterprises increasingly require infrastructure that is designed not just for experimentation, but for long-term operational scalability.

In practice, organizations are no longer managing a single model or workflow. Production AI environments often involve multiple models, varying workload types, hybrid deployment architectures, and continuously evolving cost-performance tradeoffs. This introduces a new operational challenge: orchestrating AI systems efficiently across compute, inference, storage, and deployment layers.

This is where AI-native cloud platforms are beginning to play a more important role.

Platforms such as Bitdeer AI are designed to help enterprises address these emerging deployment requirements through integrated GPU infrastructure, scalable inference environments, and flexible model deployment workflows. 

Rather than treating AI as a standalone application layer, the focus shifts toward enabling production-ready AI systems that can adapt to changing workload demands, cost constraints, and deployment strategies.

For example, with Bitdeer AI Cloud, businesses and AI teams can:

• Scale GPU resources flexibly for training, inference, embeddings, and automation workloads as demand changes.
• Support different deployment needs across cloud, dedicated infrastructure, and hybrid environments, with options such as bare metal, VM-based GPU services, and containerized workspaces.
• Improve infrastructure efficiency as token usage and inference demand grows, with better visibility into usage, costs, budgets, and resource allocation.
• Access and deploy open-source models more easily through flexible model workflows, API-based access, and integrated compute environments.
• Build AI agent workflows that connect business data, APIs, MCP tools, knowledge bases, and operational rules to support use cases such as customer support, document processing, reporting, knowledge search, and workflow automation.
• Operate AI workloads with stronger control, including managed databases, private network isolation, access controls, monitoring, and role-based management

In this way, Bitdeer AI Cloud helps businesses move beyond raw GPU capacity toward a more complete AI cloud environment — one that supports model development, deployment, and workflow automation from experimentation to production.

Conclusion

The evolution of AI has reached a point where access to models is no longer the differentiating factor. The competitive landscape is now defined by an organization’s ability to deploy, scale, and optimize AI effectively.

Success will depend on the ability to align organizational incentives, manage cost structures, and build systems that can evolve alongside rapidly advancing technology.

In this new phase, the question is not who has the best model.It is who can make AI work in the real world.

SINGAPORE, 26 June, 2026 — Bitdeer AI, part of Bitdeer Technologies Group (NASDAQ: BTDR) and a preferred NVIDIA Cloud Partner, today announced it has been the winner of “AI Cloud Platform of the Year” in 2026 AI Breakthrough Awards. The award is presented by AI Breakthrough, a leading market intelligence organization that recognizes the top companies, technologies and products in the global AI market today. The prestigious accolade recognizes Bitdeer AI’s architectural breakthrough in delivering a fully integrated, full-stack AI cloud environment optimized for enterprise-scale generative AI and production workloads.

As enterprises transition from localized AI experimentation to global production, traditional cloud architectures are fracturing under the weight of fragmented workflows, capacity shortages, and volatile pricing. Bitdeer AI solves these structural challenges through a fundamentally different approach: a vertically integrated "AI Factory" model. By owning and operating its high-performance data center infrastructure, Bitdeer AI eliminates third-party dependencies, optimizing performance directly from the physical facility and silicon level up through the core application software.

This ground-up ownership unlocks a highly unified, high-velocity software ecosystem. The Bitdeer AI cloud platform removes the friction of traditional deployment by bridging the entire AI lifecycle into a singular environment, spanning distributed training, multi-cluster scheduling, an expansive optimized model library, and serverless inference APIs.

By operating a completely integrated stack, Bitdeer AI delivers distinct enterprise advantages that redefine the AI cloud category:

• Seamless Workflow Velocity: Eliminates the need to stitch together fragmented services, allowing organizations to move from raw data to fine-tuning and sustained inference within a single, cohesive platform.
• Uncompromised Flexibility & Control: Grants developers granular control over their compute environments, offering a seamless choice between Bare Metal instances for maximum raw performance or Virtual Machines (VMs) for rapid, elastic scaling.
• Structural Cost Stability: By eliminating the margins associated with third-party infrastructure hosting, Bitdeer AI passes unprecedented price-performance leadership and transparent cost structures directly to users.
• Deep Silicon Optimization: As a preferred NVIDIA Cloud Partner, Bitdeer AI deeply integrates NVIDIA’s hardware and enterprise software stack. Controlling the physical data center allows Bitdeer AI to optimize power density and cooling specifically for next-generation NVIDIA architectures, granting clients guaranteed capacity and day-0 access to the latest software and model capabilities.

"Bitdeer AI has built one of the most complete and capable full-stack AI cloud platforms available today, combining high-performance GPU infrastructure with an enterprise-ready AI development environment," said Steve Johansson, Managing Director, AI Breakthrough. 

"We are honored to receive this recognition from AI Breakthrough for the second consecutive year," said Retainna Lin, VP of Bitdeer AI Cloud. "The winners of the next era of AI will be determined by execution speed and infrastructure reliability. At Bitdeer AI, we have engineered a full-stack cloud platform that compresses the journey from raw GPU compute to live production. We are building the foundational operating layer for global AI deployment."

Looking ahead, as the AI landscape shifts from experimentation into a mature phase defined by global production, Bitdeer AI is positioned to serve as the industry's foundational architecture. Moving forward, the company is executing on a multi-phase strategy to scale its unified cloud ecosystem globally—expanding its footprint of sustainable, proprietary "AI Factories" to meet regional growth and strict data residency requirements, deepening its integration with NVIDIA to support next-generation architectures alongside day-0 model and software releases, and continuously evolving its platform features into the definitive, full-stack operating layer for enterprise AI at scale.

Forward-Looking Statements

This press release may contain forward-looking statements regarding Bitdeer AI’s anticipated future performance, market opportunities, and business strategies. These statements are based on current beliefs, assumptions, and expectations, and are subject to risks and uncertainties that could cause actual results to differ materially. Bitdeer AI disclaims any obligation to update or revise these forward-looking statements to reflect future events or developments, except as required by law.

As autonomous agents take on increasingly complex, long-running tasks, from orchestrating multi-week coding projects to synthesizing hundreds of research sources in real time, the demands placed on the underlying reasoning model have fundamentally changed. Speed, depth of reasoning, and the ability to sustain context across extended sessions are no longer optional. They are the baseline for any model serious about agentic work.

Today, we are delighted to announce that NVIDIA Nemotron™ 3 Ultra is available at launch on Bitdeer AI Model Studio. As the flagship of the NVIDIA Nemotron family, Nemotron 3 Ultra is an open frontier reasoning model purpose-built or long-running autonomous agents. Designed to be smaller, faster, and lower cost for agent workflows, it delivers up to 5x faster inference and up to 30% lower cost while maintaining frontier-level reasoning capabilities for coding, deep research, and enterprise automation.

What is NVIDIA Nemotron 3 Ultra?

NVIDIA Nemotron 3 Ultra is an open frontier-reasoning model built for long-running autonomous agents. It is optimized for agent orchestration, complex reasoning, coding, and deep research workloads where speed, cost efficiency, and sustained reasoning depth matter as much as raw intelligence. Unlike traditional chat-focused models Nemotron 3 Ultra is designed for workflows that span hundreds of turns, multiple tools, and extended executive cycles—delivering up to 5x faster inference and up to 30% lower cost for agentic workloads

Post-trained for agent harnesses, Nemotron 3 Ultra sustains reasoning depth across the hardest calls — architectural decisions across week-long autonomous coding sessions, synthesis across hundreds of contradictory research sources, or verification of chip designs across thousands of interdependent constraints.

Fully open, Nemotron 3 Ultra can be fine-tuned for any domain and deployed on any infrastructure, giving enterprises the flexibility to maintain data control while benefiting from frontier-class intelligence.

Key Specifications

Why Faster and Lower-Cost Agent Reasoning Matters?

Most production agentic systems are not bottlenecked by a single model call — they are bottlenecked by the cumulative cost of many. A coding agent planning a complex refactor, a research agent cross-referencing hundreds of sources, or an enterprise agent triaging thousands of alerts all depend on how quickly the model can complete each reasoning cycle. Throughput is not just a hardware metric; it is a direct multiplier on how much an agent can accomplish within any given time budget.

Nemotron 3 Ultra addresses this at the architectural level:

Fastest Task Completion. Ultra's Hybrid Mamba-Transformer MoE architecture delivers the highest token throughput in NVIDIA’s published comparisons against leading open frontier baselines, enabling more reasoning cycles per time budget. Multi-Token Prediction (MTP) further reduces generation time for long sequences by predicting multiple future tokens in a single forward pass, and NVFP4 precision, optimized specifically for NVIDIA Blackwell GPUs, delivers significant inference speedup versus FP8 while maintaining accuracy.

Leading Accuracy. Latent MoE enables Ultra to call four experts for the inference cost of just one, improving intelligence and generalization with no added compute cost. Multi-environment reinforcement learning training across a broad set of agentic environments gives the model robust tool calling, reasoning, and instruction-following capabilities. The 1M token context window retains conversation history and plan states across long-running agent sessions, and enables cross-document reasoning at a scale that shorter-context models cannot match.

Fully Open. Ultra is released with open weights under NVIDIA's open-model license, trained on NVIDIA-generated high-quality synthetic data that is fully open, and accompanied by published development techniques and recipes, giving researchers and enterprises full transparency and the flexibility to customize or build on top of the model.

Enterprise Use Cases

Source: NVIDIA

Programming and Coding Agents. Coding agents built on Ultra plan, code, test, debug, and iterate end-to-end across large codebases. Ultra handles the hard reasoning calls: architectural planning, complex multi-file refactors, and error recovery, sustaining coherent reasoning across sessions that can span days or weeks.

Deep Research and Search. Research agents search, evaluate, cross-reference, and synthesize across hundreds of sources in sustained parallel loops. Ultra handles final synthesis, the step where contradictions must be resolved, gaps must be identified, and novel hypotheses proposed with the depth and consistency that frontier reasoning demands.

Enterprise Workflow Agents. Agents built for enterprise workflows automate operations across industries in persistent, tool-using loops: triaging thousands of security alerts, ingesting and interpreting regulatory filings, orchestrating clinical trial operations. Ultra handles the complex reasoning steps within these workflows, where errors in judgment have real downstream consequences.

EDA and Chip Design. Chip design agents autonomously generate RTL from specifications, verify designs across thousands of constraints, and orchestrate workflows from design to manufacturing sign-off. Ultra handles verification, failure analysis, and cross-block dependency resolution — the reasoning-intensive operations that define the quality of the final design.

Supported Agentic Frameworks

Nemotron 3 Ultra integrates with leading open agent frameworks out of the box — from single-command deployment with NVIDIA NemoClaw to tested cookbooks for the most popular coding and orchestration platforms. It is also packaged as an NVIDIA NIM microservice, making it deployable across data center and cloud environments without modification.

Run Nemotron 3 Ultra Via API on Bitdeer AI Model Studio

You can run Nemotron 3 Ultra on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable.With a unified API, Model Studio allows developers and enterprises to start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value.

Bitdeer AI is a preferred NVIDIA Cloud Partner, certified to ISO/IEC 27001:2022 and SOC2 Type I & Type II, providing the secure, compliant, high-performance, and enterprise-grade infrastructure that production agentic AI deployments require. Run your models at the precision and scale your business requires, on Bitdeer AI's purpose-built GPU fleet.

Get Started

1. Log in to Bitdeer AI Model Studio
2. Locate NVIDIA Nemotron 3 Ultra in the model list
3. Generate an API key and start making API calls

This streamlined workflow enables rapid integration of frontier reasoning capabilities into applications and agent systems.

Conclusion

NVIDIA Nemotron 3 Ultra raises the bar for what open frontier models can deliver in production. Combining frontier-class intelligence with up to 5x faster inference and up to 30% lower cost, Ultra is designed specifically for the demands of long-running autonomous agents. With a 1M - token context window, open weights, open training recipes, and  With day-0 availability on Bitdeer AI Model Studio, organizations can move quickly from experimentation to production while maintaining performance, flexibility, and economics required for large-scale agentic AI.

As AI agents evolve beyond text and into real-world workflows, the ability to understand information and reason across multiple modalities is becoming essential. From video and audio to documents and UI screens, modern agentic systems require models that can reason across modalities with accuracy, efficiency, and enterprise readiness. 

Today, we’re delighted to announce that NVIDIA Nemotron™ 3 Nano Omni is available at launch on our Bitdeer AI Model Studio. As part of the broader NVIDIA Nemotron family, it represents a new step forward in open, production-ready multimodal reasoning models.

What is NVIDIA Nemotron 3 Nano Omni

NVIDIA Nemotron 3 Nano Omni is an open multimodal foundation model designed for production agentic AI. It unifies reasoning across video, audio, images, documents, charts, and text in a single model—eliminating the need for fragmented multimodal pipelines.

By consolidating perception and reasoning into one system, Nemotron 3 Nano Omni simplifies agent development, reduces orchestration complexity, improves efficiency and scalability, and delivers leading multimodal accuracy. Built with open weights, datasets, and recipes, it enables developers and enterprises to customize, deploy, and operate multimodal agents with full control and flexibility.

Key Specifications:

●  Model Size: 30B A3B

●  Modalities: Input (text, image, video, audio), Output (text).

●  Architecture: Hybrid Mixture of Experts (MoE) with Transformer-Mamba design

●  VIsion Encoder: CRADIO v4-H

●  Audio ENcoder: Parakeet

●  Context Length: 256k

●      Optimizations:

• Conv3D fortemporal video reasoning
• Efficient Video Sampling (EVS) for lower inference cost

●      Quantization: FP8, NVFP4

Why a Unified Multimodal Model Matters

Many enterprise AI systems still rely on stitched-together pipelines across vision, speech, OCR, and reasoning models. This approach introduces higher latency from repeated inference passes, increased operational complexity, and fragmented context across modalities.

Nemotron 3 Nano Omni addresses this by acting as a multimodal perception and reasoning layer within agent systems—enabling a unified perception → reasoning → action loop.

Enterprise use cases

Customer Service Agent: A customer service agent operates in a highly multimodal environment. It analyzes recorded customer interactions, including audio and speech transcriptions. It reasons over screen recordings of customer sessions and images such as screenshots of errors or invoices. At the same time, it reads documents like knowledge‑base articles, policies, and CRM history. Nemotron 3 Nano Omni unifies all of these signals so the agent understands not just what the customer said, but what they experienced and what the business rules allow—enabling accurate, context‑aware resolution.

Financial Analyst Agent: Financial analysis depends on more than text alone. This agent reasons across documents like financial filings and earnings transcripts, images such as charts and scanned reports, audio and speech from earnings calls, and video from investor presentations. Nemotron 3 Nano Omni ties together what executives say, how numbers are presented visually, and what the underlying documents show—producing grounded insights rather than surface‑level summaries.

Computer Use Agent: The computer use agent is one of the clearest demonstrations of unified multimodality. It analyzes video and images from screen recordings to understand UI state over time, interprets instructions and system audio cues, and reads documents like task instructions and validation policies. Nemotron 3 Nano Omni enables the agent to see the interface, understand intent, read constraints, and take the correct action—all within one reasoning loop. This collapses when perception and decision‑making are split across models.

Media and Entertainment Agent: Media workflows depend on more than transcripts alone. This agent reasons across video content, dialogue, on-screen text, and visual scene changes to support richer video and speech analysis. Nemotron 3 Nano Omni can generate dense captions that capture not just what is said, but what appears and happens on screen, while also improving video search and summarization across large content libraries. This helps media teams turn raw footage into searchable, contextualized, and production-ready assets more efficiently.

Run Nemotron 3 Nano Omni Via API on Bitdeer AI Model Studio

You can run Nemotron 3 Nano Omni on Bitdeer AI Model Studio, our serverless inference platform designed to make access to advanced foundation models simple and scalable. With a straightforward API, Model Studio allows developers and enterprises to start using models quickly without managing underlying infrastructure, reducing deployment complexity and time to value.

This makes it easier to integrate multimodal reasoning capabilities into applications and agentic workflows, while benefiting from a more flexible and efficient path from experimentation to production in just a few steps.

Get Started

1. Log in to Bitdeer AI Model Studio
2. Locate Nemotron 3 Nano Omni in the model list
3. Generate API key and start making API calls

This streamlined workflow enables rapid integration of multimodal reasoning into applications and agent systems.

Conclusion

NVIDIA Nemotron 3 Nano Omni makes multimodal AI more practical for real-world deployments. Instead of managing multiple models and pipelines, teams can focus on building agentic applications, automating workflows, and delivering better user experiences.

With availability on Bitdeer AI Model Studio, organizations can move from experiment to production faster—turning multimodal AI into measurable business impact.

OpenClaw has surpassed 290,000 stars on GitHub, making it one of the most popular open-source AI Agent frameworks of 2026. A growing number of developers and tech enthusiasts are experimenting with deploying their own AI assistants locally.

However, local deployment presents significant challenges in practice: service interruptions when the machine shuts down, complex environment configuration, API key security risks, and more — all of which severely limit the continuous availability of an AI Agent.

Deploying OpenClaw to the cloud addresses these issues at their root. This article examines the key advantages of cloud deployment, provides guidance on selecting the right large language model, and demonstrates real-world use cases for OpenClaw.

What Is OpenClaw?

OpenClaw is an open-source personal AI Agent framework built with TypeScript. Its core capabilities include:

• 24/7 Autonomous Operation: A built-in heartbeat mechanism proactively monitors tasks and takes action without requiring continuous user intervention
• 50+ Platform Integrations: Supports WhatsApp, Telegram, Discord, Slack, Gmail, GitHub, and other major platforms for unified cross-platform management
• Model Flexibility: Compatible with cloud-based LLMs such as Claude, GPT, and Grok, as well as local models via Ollama / vLLM
• Extensible Skills System: Skills are defined using Markdown + YAML, with 100+ pre-built skills available on ClawHub
• Privacy-First Design: All data is stored locally by default, with the memory system based on local Markdown files

Unlike traditional chatbots, OpenClaw features contextual memory, proactive behavior, and tool invocation capabilities — functioning more as a persistent AI collaborator than a simple Q&A interface.

Cloud Deployment vs. Local Deployment

While OpenClaw supports local installation, cloud deployment offers clear advantages for users who require stability and continuous availability.

Local deployment is suitable for short-term testing, while cloud deployment is what enables an AI Agent to deliver continuous, reliable service. An assistant that needs to be always available requires an environment that is always online.

Deployment Cost

Running OpenClaw does not require GPU resources. A basic CPU virtual machine instance (2 cores / 4GB RAM / 20GB SSD) is sufficient for stable operation. At this configuration, Bitdeer AI Cloud's on-demand pricing is approximately $0.0363/hour (around $26/month) — providing a 24/7 online AI assistant environment at a fraction of typical infrastructure costs.

Bitdeer AI Cloud, as an NVIDIA Preferred Cloud Service Provider, offers compute resources spanning multiple generations of NVIDIA GPUs including H100, H200, B200, and GB200. For Agent deployment scenarios like OpenClaw, a CPU-only instance is all that's needed. For advanced use cases such as model fine-tuning or local inference, GPU instances are available for seamless scaling. Virtual machine instances are billed on-demand — for detailed pricing, refer to the GPU Compute Pricing page.

Model inference is handled via API calls and billed per token. Bitdeer AI Cloud provides multi-vendor inference services covering text generation and other task types. For detailed inference pricing, refer to the AI Models Pricing page.

Model Selection Guide

One of OpenClaw's core strengths is model flexibility — it is not locked into any single LLM provider. Bitdeer AI Cloud currently offers 40+ models from 10+ providers, spanning text generation, visual understanding, reasoning, and image generation. All APIs are OpenAI-compatible, making model switching nearly effortless.

Below are representative text generation models suitable for OpenClaw, with pricing reference:

Scenario-Based Recommendations

• Best overall value — DeepSeek-V3.2, with the lowest combined input/output cost among flagship models and industry-leading code and reasoning capabilities
• Long-context input scenarios (e.g., document analysis, email summarization) — Qwen3-235B-A22B, with input pricing as low as $0.22 per million tokens
• Deep conversational AI — Kimi-K2.5, with excellent language comprehension for high-quality conversational output
• Ultra-low cost — The platform also offers lightweight models such as Gemma and Ministral, with input pricing starting at $0.02 per million tokens for high-frequency, low-complexity tasks

For the full model catalog and interactive demos, visit Bitdeer Model Explore.

OpenClaw integrates with Bitdeer's model service entirely through its configuration file — users simply enter their API key and model name in openclaw.json, and OpenClaw handles all API communication automatically with no additional code required. All Bitdeer model APIs are OpenAI REST-compatible, allowing OpenClaw to recognize and invoke them directly.

API keys can be generated in the Bitdeer AI Cloud console under API Keys.

Deployment Guide

Bitdeer AI has published a comprehensive deployment tutorial covering everything from creating a cloud instance and installing OpenClaw to configuring model connections and integrating with Telegram. For the full step-by-step guide, refer to:

Installing and Configuring OpenClaw on Bitdeer AI Cloud

The deployment process can be summarized in five steps:

1. Create a Cloud Instance — Provision a virtual machine on Bitdeer AI Cloud
2. Install OpenClaw — Run the official one-line installation script
3. Configure Models — Enter your Bitdeer AI API key and specify the LLM
4. Connect Telegram — Pair with a Telegram Bot for mobile access
5. Run in Background — The service runs continuously 24/7 with no manual maintenance required

Real-World Use Cases

Once deployed, OpenClaw delivers practical value across multiple dimensions. Below are several representative use cases:

Personal Productivity

• Connect to Gmail to automatically read emails, extract action items, and send daily digests on a set schedule
• Integrate with Google Calendar to create and manage events directly through conversation
• Perform real-time multilingual translation across English, Chinese, Japanese, Korean, and more

Research and Analysis

• Execute web searches, read page content, and generate structured summaries
• Extract key points and summaries from any given URL
• Conduct multi-step research with comparative analysis and structured report output

Code Generation and Execution

• Generate and execute code in a sandboxed environment, returning results directly
• Support for Python, JavaScript, and other mainstream languages for data processing and scripting
• Combine web search capabilities with code generation to automatically reference documentation and produce runnable solutions

These capabilities are powered by OpenClaw's skills system. Over 100 pre-built skills are available on ClawHub for immediate use. For custom skill development, refer to the OpenClaw Skills Documentation.

Conclusion

As a fully featured open-source AI Agent framework, OpenClaw — combined with Bitdeer AI Cloud's infrastructure and model inference services — provides a stable, secure, and always-online AI assistant solution.

Get Staarted

• Sign up for Bitdeer AI Cloud
• View the full deployment tutorial
• Visit the OpenClaw GitHub repository
• Bitdeer AI Documentation Center

Note:Pricing for models and GPU resources is subject to change. Please refer to the platform for the most up-to-date pricing.