Introducing Typhoon-S: A Blueprint for Sovereign AI

Over the past few years, the conversation around AI has shifted. It is no longer just about who can build the biggest model—but about who controls AI systems, how they are trained, and whether they truly serve local needs.

For many countries, institutions, and regulated domains, reliance on opaque, foreign-built models is increasingly untenable. Cost, data governance, cultural alignment, and legal accountability all point toward the same conclusion: AI must become sovereign—understandable, controllable, and adaptable by those who deploy it.

At Typhoon, this belief has guided our work since day one.

It has been over two years since Typhoon was founded, with a simple goal: making Thailand a better place through AI research. In 2024, we released a series of research previews - from Typhoon Vision and Audio to Typhoon T1, R1, and Typhoon 2. These were among the first models of their kind in Southeast Asia, proving that we can produce world-class AI research right here. While our mandate has evolved toward more task-specific models like Typhoon OCR, Translate, and ASR, our passion for advance the Thailand’s place on the global AI stage is defined by our commitment to Sovereignty.

We aren't just building models; we are advocating for a world where AI is localized and controllable. We are dedicated to share and publishing open research and technical report that help for the global community to build sovereign systems that are precise, and cultural align while resource-efficient.

Today, we are excited to release Typhoon-S (Sovereign). This isn't just another model; it is our blueprint for building high-performing, region- and domain-specific LLMs under strict resource constraints—proving that world-class AI research doesn't require a trillion-dollar budget.

TL;DR: We are releasing everything.

To dismantle resource gatekeeping and empower the global community, we are fully open-sourcing the Typhoon-S project.

• 📜 Technical Report:
  A comprehensive report of our minimal post-training recipe (SFT + OPD) and our InK-GRPO method for domain specialization.
• 💻 Code:
  Our reference implementation for training pipelines: general on-policy logits distillation with dynamic model swapping and InK-GRPO for sovereign and long-tail domain adaptation.
• 📊 Datasets:
  • The Typhoon-S-Instruct-Dataset, including our Thai AutoIF and cross-lingual alignment data, and the Typhoon-S-Sovereign-Dataset, used to evaluate and train for specific Sovereign Capabilities.
• 🤖 Models:
  • The Typhoon-S-ThaiLLM-8B-Instruct, built on the sovereignty-focused ThaiLLM-8B model and shown to outperform global baselines on native Thai tasks.
  • The Typhoon-S-4B-Legal-Agent, demonstrating that domain-specific sovereignty can outperform brute-force scale.

The Problem: Resource Gatekeeping

Today’s state-of-the-art LLMs are developed by a small number of organizations with access to massive compute budgets and globally scraped, English- and Chinese-centric data.

This concentration creates a form of resource gatekeeping. Even when models are labeled “open,” their training recipes often depend on expensive pipelines, proprietary data mixtures, or reinforcement learning setups that are inaccessible to most teams.

For Sovereign AI—systems where a country, institution, or domain owner must retain control over model weights, data, and deployment—this is a fundamental barrier.

In practice, sovereign initiatives face two competing needs:

1. Adoptability

The ability to turn a base model into a useful general assistant with strong instruction-following, reasoning, and tool-use—without billion-dollar compute.

1. Sovereign Capability

The ability to perform high-stakes, region- or domain-specific tasks (such as legal reasoning or culturally grounded logic) that are underrepresented in global datasets.

Making Sovereign Models Useful (Base → Instruct)

A central challenge for Sovereign AI is adoptability.

How can a regional or national initiative take a base model and transform it into a capable assistant—without relying on the same expensive post-training pipelines used by frontier labs?

We demonstrated that with a sovereignty-focused foundation—ThaiLLM-8B, a base model continued-pretrained on 64B Thai tokens. By applying our minimal (supervise finetuning) SFT + (on-policy distillation) OPD recipe, we can create a model that remains strong on native Thai usage while gaining general assistant behavior.

Sovereign vs Global Model Performance on Native Thai Benchmarks

This table compares Typhoon-S-8B, a sovereignty-focused model built on ThaiLLM-8B, against Qwen3-8B, a strong global baseline, across a suite of native Thai benchmarks written originally in Thai rather than translated.

Sovereignty as a Performance Advantage

Global models excel at multilingual breadth, but sovereign-adapted models begin with depth: regional language patterns, cultural context, and localized reasoning.

By combining a locally grounded base model with efficient post-training, Typhoon-S demonstrates that sovereign models can achieve competitive general assistant behavior and superior performance on native tasks.

On our fully Thai benchmark suite, Typhoon-S-8B achieved a significantly higher average score than Qwen3-8B. This demonstrates that Sovereign Capability isn't just about control—it's about better performance for the people who actually use the language.

Pushing the Frontier: Sovereign Capability

General-purpose assistants are useful—but Sovereign AI must go further.

High-stakes domains like law, medicine, and public policy require models that understand local institutions, documents, and interpretive norms. These domains are often poorly represented in large, general-purpose datasets.

To address this, we explored methods that move beyond standard instruction tuning.

The Innovation: InK-GRPO

Standard Reinforcement Fine-Tuning (RFT) is excellent at "amplifying" reasoning patterns, but research suggests it rarely introduces new factual knowledge. For a sovereign setting like Thai Law, if the knowledge isn't in the base model, RL alone won't find it.

We introduce InK-GRPO (Injected Knowledge GRPO).

This approach combines the reasoning power of the GRPO algorithm with a stochastic "Knowledge Injection" step.

• How it works: During the reinforcement learning process, we periodically inject a next-token prediction objective (Cross-Entropy loss) using in-domain text (e.g., Thai legal documents).
• The result: The model doesn't just learn how to reason; it learns the content it needs to reason about.

We put InK-GRPO to the test on NitiBench, a challenging benchmark for Thai legal reasoning. We found that jointly optimizing for domain knowledge during RL significantly boosted accuracy compared to standard RL alone.

Agentic Sovereignty: When 4B Beats Frontier Models

Sovereign tasks often require more than a single response; especially in high-stakes fields like Law or Medicine. You don't just want a model that "remembers" a law; you want an Agent that can retrieve regulations, cross-reference documents, and reason correctly under local legal interpretation.

To achieve this, we moved beyond static Q&A. We placed our models into a controlled Retrieval-Augmented Generation (RAG) environment, equipping them with search and read tools to interact with a Thai legal database.

We then applied Agentic RFT (Reinforcement Fine-Tuning), where the model was rewarded not just for the generating the answer, but for its ability to navigate external knowledge effectively across multiple turns.

The results were startling. By specializing a 4B-parameter model through Agentic RFT, it reached a level of legal accuracy that surpassed much larger, general-purpose "GPT-5" level baselines in the same environment.

This proves that domain-specific sovereignty can outperform brute-force scale. We’ve shown that a focused 4B model, empowered with the right tools and training, is more capable for a nation’s specific needs than the world’s largest general-purpose models.

Develop the Future of Sovereign AI Together

We are fully opening the recipe, technical report, datasets, and models to help others build their own sovereign LLMs.

• 📜 Technical Report:
  A comprehensive report of our minimal post-training recipe (SFT + OPD) and our InK-GRPO method for domain specialization.
• 💻 Code:
  Our reference implementation for training pipelines: general on-policy logits distillation with dynamic model swapping and InK-GRPO for sovereign and long-tail domain adaptation.
• 📊 Datasets:
  • The Typhoon-S-Instruct-Dataset, including our Thai AutoIF and cross-lingual alignment data, and the Typhoon-S-Sovereign-Dataset, used to evaluate and train for specific Sovereign Capabilities.
• 🤖 Models:
  • The Typhoon-S-ThaiLLM-8B-Instruct, built on the sovereignty-focused ThaiLLM-8B model and shown to outperform global baselines on native Thai tasks.
  • The Typhoon-S-4B-Legal-Agent, demonstrating that domain-specific sovereignty can outperform brute-force scale.

Current Limitations & Future Work

• Post-training: We currently frame at only post-training stage only as pre-training and mid-training is beyond our scale.
• Scaling: Our current experiments were capped at 8xH100 GPUs. We are eager to see how these techniques scale to larger clusters.
• Generalization: While we used Thai as our primary case study, the OPD and InK-GRPO recipes are language-agnostic.

We believe that high-quality AI should not require trillion-dollar budgets. With the right foundations and recipes, sovereign institutions can build systems that are not only useful—but deeply aligned with their people, laws, and cultures.