typhoon-ai/typhoon-si-med-thinking-4b-research-preview
Typhoon-Si-Med-Thinking-4B is Southeast Asia's first 4-billion-parameter medical reasoning model, developed by Typhoon (SCB 10X) and Siriraj Informatics and Data Innovation Center (SiData+). Based on Qwen3 architecture and trained with reinforcement learning, it generates ranked lists of candidate answers for medical questions, outperforming larger models like Gemini 2.5 Pro on list-based and short-answer medical QA benchmarks. This model is optimized for efficient, domain-specific medical reasoning, supporting a 40960 token context length.
Loading preview...
Overview
Typhoon-Si-Med-Thinking-4B is a 4-billion-parameter instruction-tuned decoder-only model, jointly developed by Typhoon (SCB 10X) and the Siriraj Informatics and Data Innovation Center (SiData+). It is based on the Qwen3 architecture and utilizes reinforcement learning to generate ranked lists of candidate answers for medical reasoning tasks. This approach aims to better reflect the uncertainty in clinical decision-making compared to traditional single-answer formats.
Key Capabilities
- Ranked-List Reasoning: Generates a ranked list of plausible answers, mirroring clinical thought processes and fostering collaborative reasoning.
- Robust Performance: Achieves strong results on medical QA benchmarks including MedQA, MedMCQA, MedXpertQA, and MMLU Pro (Health).
- Efficiency: A small, efficient model that surpasses larger systems like Gemini 2.5 Pro on list-based and short-answer medical tasks.
- Dual Reasoning Modes: Supports
TEXT_MODEfor a single answer with reasoning trace andLIST_MODEfor a ranked list of answers with reasoning trace. - Clinical Assistant: Designed as a reasoning-enabled clinical assistant model, outputting both intermediate reasoning and final answers.
Intended Use & Limitations
This model is an instructional reasoning model and a research preview. It is not intended for medical use and may produce inaccurate, biased, or objectionable answers. Developers are advised to assess risks for their specific use cases. For more details, refer to the research paper.