Typhoon-S
Research
Qwen3, ThaiLLM
4B
8B
Minimal Open Post-Training for Sovereign LLM (Research Preview)
Released
January 29, 2025
Context
Input
Text
Output
Text
About this Model
Typhoon-S is an open blueprint for building high-performance sovereign LLMs, designed to align with local language, culture, and domain needs under real-world resource constraints.
Key Features
Typhoon-S-ThaiLLM-8B-Instruct: Adoptability
Transforms sovereignty-focused base model (ThaiLLM 8B) into capable general assistant using minimal post-training. Outperforms Qwen3-8B on native Thai benchmarks.
Typhoon-S-4B-Legal-Agent: Sovereign Capability
Domain-specific 4B model achieves 78.02% on NitiBench Thai legal reasoning, surpassing GPT-5 (75.34%) despite being much smaller.
Minimal Post-Training Recipe
Lightweight pipeline combining supervised fine-tuning with on-policy distillation to achieve general assistant behavior while preserving regional depth.
InK-GRPO: Knowledge Injection During RL
Novel method injecting domain knowledge during reinforcement learning, enabling models to learn both how to reason and what to reason about.
Fully Open Blueprint for Sovereign AI
Complete release: technical report, code, datasets, and models to democratize sovereign AI development.
Release History
Version 1
January 29, 2025
Initial release
Availability
Web Playground
Not available
Typhoon API
Not available
Typhoon API Pro
Not available
Other Platforms
Not available
Related Links