Open Models AI Codefest: The Era of the AI Demo Is Over. The Era of Production and Localization Is Here.

banner

Author: Angel Hsu, Evelyn Xiong


For years, much of the AI conversation has revolved around what models can do. But as AI adoption accelerates, the more important question is becoming: Can these systems work reliably, economically and securely in the real world?

That shift was evident throughout the recent Open Models AI Codefest, hosted by NVIDIA and Bitdeer AI.

Ten teams from across Southeast Asia spent several weeks tackling complex, real-world challenges spanning drug discovery, regulatory compliance, consumer behavior modeling, mortgage origination, localized language models and autonomous mobile engineering.

The goal was to push open models beyond experimentation and closer to production.And the most interesting outcome was not simply what the models could achieve.

It was what the teams learned about what actually makes AI work in practice.

Across very different industries and use cases, several common lessons emerged: model size alone is no longer the primary differentiator. Increasingly, performance comes from the quality of the data, the architecture built around the model, the ability to deploy specialized models for specific tasks, and the software and infrastructure that make rapid iteration possible.

Here is what the cohort showed us.

One Codefest. Ten Teams. Five Countries. One Region Building Forward.

The participating teams approached AI from very different directions:

  • NYB.AI explored multi-tool AI agents for drug discovery.
  • Berdaulat developed a sovereign Indonesian fintech LLM.
  • Bits2Agents worked on an intelligent mortgage origination platform.
  • ARAC built customer digital twins for behavioral simulation.
  • MISA LLM focused on continual pre-training for Vietnamese.
  • LynxKite explored generative molecular dynamics.
  • Little Fruit developed a school nutrition auditing platform.
  • Sharp Minds worked on sovereign compliance infrastructure.
  • Team A developed its Prism engine and specialized small models.
  • PANIC built an autonomous mobile engineering system.

They received support from NVIDIA and Bitdeer AI experts, backed by Bitdeer AI's AI infrastructure and technical resources, a dedicated node of eight NVIDIA H200 Tensor Core GPUs per team, together with the NVIDIA AI software stack for data curation, synthetic data generation, training, inference and agent development.

Although the projects looked very different on the surface, the engineering challenges beneath them were surprisingly similar, and so was the shape of the toolchain the teams converged on.

1. Localization Is Becoming a Competitive Advantage

Southeast Asia is one of the most linguistically diverse regions in the world, yet many frontier models still underperform in local languages or carry unnecessary complexity for highly specific regional use cases.

The Codefest challenged the assumption that solving this problem requires enormous models and equally enormous budgets.

In many cases, the opposite proved more effective: narrow and specialized can outperform broad and general-purpose.

Bits2Agents achieved a 10% Vietnamese word error rate, matching incumbent commercial cloud services, by focusing on a highly targeted bilingual model rather than relying on a broad, off-the-shelf multilingual system. The team post-trained that bilingual Vietnamese-English model with the NVIDIA NeMo framework, using its post-training recipes to concentrate model capacity on the vocabulary that mortgage conversations actually contain instead of chasing broad multilingual coverage.

Team A demonstrated a similar principle for Malay, using a mixture-of-experts approach that exceeded a regional benchmark while reducing runtime costs significantly compared with dense alternatives. The team even demonstrated a near-frontier-capable model running offline on an Android device.

The implication for Southeast Asian AI builders is important.

Localization is no longer simply about translating a global model into a local language. It is becoming an engineering discipline of its own - one built around local data, local workflows, local constraints and increasingly, smaller specialized models.

2. Data and the Harness Matter More Than Constant Model Tweaking

Perhaps the strongest common lesson across the cohort was that better AI did not necessarily come from endlessly adjusting the underlying model.

It came from improving everything around it.

Berdaulat found that its meaningful performance gains consistently came from changes to the data pipeline rather than hyperparameter or architectural tweaks. Because much of the regulated data it needed could not be used freely for training, the team generated domain-specific synthetic Indonesian financial data from a small set of seed examples with NVIDIA NeMo Data Designer, then used NVIDIA NeMo Curator to deduplicate, filter and rebalance that corpus into training mixtures. Rebuilding a mixture was cheap enough that data composition became something to experiment with rather than assume , which is how the team was able to locate its gains in the pipeline in the first place.

MISA LLM reached a similar conclusion: data quality mattered more than sheer data quantity, while the balance of the training mixture played a critical role in performance. The team ran large Vietnamese collections through NVIDIA NeMo Curator, applying exact, fuzzy and semantic deduplication, language identification and quality classification to strip out the near-duplicates and low-quality text that would otherwise have consumed training tokens , the clean baseline it needed before quality and quantity could be compared at all.

NYB.AI went even further, choosing not to fine-tune its foundation models at all.

Instead, the team focused on a provider-agnostic agent architecture capable of dynamically selecting from 311 specialized biology and chemistry tools. It started from the NVIDIA AI-Q Blueprint, NVIDIA's reference architecture for agents that connect to, retrieve from and reason over enterprise data, which gave the team a working orchestration pattern on day one rather than a blank repository, and used the NVIDIA NeMo Agent Toolkit to instrument the workflow end to end , exposing where in that 311-tool selection path latency and cost were accumulating, the visibility that made iterative tuning of the harness possible at all. The NVIDIA NeMo framework was reserved for benchmarking and customizing the smaller specialist models in the pipeline, while the foundation models were deliberately left untouched. Its conclusion was that the ultimate ceiling of the system depended less on the underlying model than on the quality of the agent and harness logic surrounding it.

This signals a broader change in AI engineering.

The competitive advantage is increasingly moving from simply having access to a powerful model to knowing how to orchestrate models, tools, data and workflows around a specific outcome.

3. Compute Changes How Teams Experiment

Moving from an idea to a production-ready system requires more than good models and strong engineering.

It also requires the freedom to experiment.

Access to Bitdeer AI's AI infrastructure powered by NVIDIA H200, together with NVIDIA's software ecosystem, allowed teams to iterate at a pace that would have been difficult under constrained compute environments.

For LynxKite, molecular training trajectories could take roughly 30 hours per run. High-performance GPU capacity allowed the team to work through failed experiments, refine its approach and complete a working generative model within the program.

For ARAC, GPU infrastructure supported a demanding four-stage NVIDIA NeMo framework training pipeline for its behavioral digital-twin system, ultimately producing a model that the team reported could outperform frontier alternatives at significantly lower cost. The framework's distributed training and post-training stages let the team fit all four stages onto its GPU node rather than trading one away.

For MISA LLM, continual pre-training ran on the NVIDIA NeMo framework, whose parallelism strategies kept the GPUs saturated from run to run , which is what made it practical to sweep several training mixtures instead of committing early to one.

For Berdaulat, access to a node of NVIDIA H200 GPUs meant the team could repeatedly test different data mixtures rather than making assumptions about which combinations would perform best. Fine-tuning ran on NVIDIA NeMo AutoModel, whose Hugging Face-native training path let open checkpoints go straight into training with no conversion work between iterations, so testing a new mixture cost the team a training run rather than a rebuild.

That difference matters.

When compute is scarce, engineering teams optimize around scarcity.

When sufficient compute is available, they can instead optimize around the problem.

4. Compute Abundance Can Become a Data Strategy

Several teams also used compute provided by Bitdeer AI in a less obvious way: to create the data they needed.

Instead of treating limited datasets as a fixed constraint, teams such as LynxKite, Little Fruit and ARAC used AI infrastructure to simulate environments and generate synthetic training data.

These ranged from molecular trajectories to food-tray imagery and behavioral simulations.

NVIDIA tools also played an important role at every stage of that pipeline.

NVIDIA NIM helped teams such as Little Fruit and ARAC support production inference with sufficiently low latency for interactive workloads , for ARAC, keeping response times inside the interactive budget a simulation tool needs, so a digital twin could be queried like a product rather than run like a batch job.

Upstream of that, NVIDIA NeMo Data Designer gave teams a declarative way to design domain-specific synthetic datasets from scratch or from seed examples. It is what turned ARAC's scarce, sensitive customer-interaction data from a hard constraint into a design variable, and what allowed Berdaulat to assemble an Indonesian financial corpus it could not have collected directly. NVIDIA NeMo Curator then turned raw, mixed-quality regional corpora into training-ready mixtures through GPU-accelerated filtering, classification and deduplication, and open NVIDIA Nemotron models, datasets and recipes gave several teams a starting point that included the data and the training recipe, not only the weights.

The lesson is larger than the individual tools.

As AI development matures, infrastructure is not simply where a finished model runs.

It increasingly influences how the model is trained, tested, evaluated, and improved.

5. Sovereign AI Is Becoming a Product Requirement

For several teams, self-hosting was not a technical preference.

It was central to whether the product could exist at all.

In financial services, compliance and other regulated industries, sensitive data often cannot be sent freely to external APIs.

Bits2Agents migrated a mortgage-origination workflow, including KYC, OCR and fraud-detection components, to self-hosted infrastructure so regulated lending data could remain within approved environments. Its speech model runs as a self-hosted speech NVIDIA NIM microservice on the NVIDIA Riva speech AI runtime, which supplied the streaming recognition pipeline and concurrency handling the team would otherwise have had to build itself , the practical difference between a compliant architecture and a compliant intention.

Berdaulat similarly designed its fintech LLM to operate in an air-gapped environment, enabling financial reasoning without relying on public model APIs , the customized model packaged as an NVIDIA NIM microservice and served with NVIDIA Dynamo behind it, disaggregating prefill and decode and routing requests with KV cache awareness so an air-gapped deployment could still hold latency steady as concurrency grew.

Sharp Minds built its retrieval and reasoning architecture around self-hosted infrastructure because data privacy was not merely an infrastructure consideration, it was part of the product promise itself.

Open models are what make this posture practical. When weights, datasets and training recipes are open, a regional team can inspect a model, retrain it on its own language and domain data, and deploy it wherever regulation requires , decisions a closed API leaves to somebody else. That is the premise behind NVIDIA Nemotron, a family of open multimodal models published with open weights, open datasets and open training recipes under a permissive license that lets teams use, modify and commercially deploy the models and their derivatives. Around those models, NVIDIA also publishes blueprints, development toolkits and data strategies , NeMo Curator and NeMo Data Designer for the dataset, NeMo Framework and NeMo AutoModel for customization, NIM, Dynamo and the AI-Q Blueprint for deployment , so a team's starting point is a reference design it can take apart and rebuild for its own market rather than an empty repository. Several Codefest teams reached a working system in weeks for exactly that reason.

This represents an important evolution in the sovereign AI conversation.

Data residency and sovereignty are increasingly becoming architecture decisions made from day one, rather than compliance requirements added after deployment.

For Southeast Asia, where regulatory environments, languages and enterprise requirements differ considerably across markets, that distinction is especially important.

6. The Agent Is Becoming More Than a Chat Interface

Another clear signal from the Codefest was the shift away from simple conversational AI.

These teams were not primarily building chatbots.

They were building systems that could reason, act, plan, simulate and use tools.

NYB.AI created an agent capable of navigating more than 300 specialized biology and chemistry tools. NVIDIA NeMo Retriever turned scientific literature and internal documents into a structured knowledge layer the agent could ground its answers in, NVIDIA NeMo Guardrails applied programmable rails to its tool calls and outputs , which matters in a scientific domain where a confidently wrong answer is expensive , and the NVIDIA NIM Operator managed the lifecycle of the underlying NIM microservices on Kubernetes as the tool count grew.

ARAC developed a digital twin that does not simply respond to questions, but simulates customer behavior to help businesses test commercial strategies.

PANIC explored autonomous mobile engineering workflows.

Across these projects, the interface became less important than the system behind it.

The future of enterprise AI is increasingly about orchestrating models and tools toward an outcome, not simply generating a better answer in a chat window.

A Blueprint for Southeast Asia's AI Future

The most important lesson from the Codefest may be that the barriers to building sophisticated AI systems are changing.

Open models have dramatically expanded access to AI capability.

But access to models alone is not enough.

The teams showed that localized, production-ready AI increasingly depends on four things: high-quality data, intelligent orchestration, specialized models and the software and infrastructure capable of supporting rapid experimentation and deployment.

Together, these elements change what is possible for AI-native companies in Southeast Asia.

A team building for Vietnamese speech no longer has to accept a global model optimized for dozens of languages.

A financial institution does not necessarily have to send sensitive data to an external model API.

A drug-discovery startup can give an AI agent hundreds of scientific tools and allow it to decide which ones to use.

And an engineering team can increasingly build systems that would once have required the resources of a much larger organization.

That is where the combination of Nemotron open models, NVIDIA's software ecosystem and Bitdeer AI's high-performance AI infrastructure becomes meaningful.

Infrastructure is not innovation by itself.

It gives innovators room to experiment, localize, iterate and ultimately bring AI into production.

The next wave of transformative AI in Southeast Asia will not simply be imported from elsewhere.

It will be built around the region's own languages, industries, regulatory environments and problems.

It will be more specialized.

More sovereign.

More production-ready.

And increasingly, it will be built right here in the region.

In the coming weeks, we'll take a closer look at individual use cases from the Codefest, exploring how these teams turned Nemotron open models, NVIDIA software, infrastructure and local expertise into real-world AI applications.