typhoon-ai/typhoon2.5-qwen3-4b

Typhoon2.5-Qwen3-4B: Thai-English Instruction Model

Typhoon2.5-Qwen3-4B is a 4 billion parameter instruction-tuned large language model developed by SCB 10X, built upon the Qwen3 architecture. It is distinguished by its extensive 256K context length and integrated function-calling capabilities, making it suitable for complex, multi-turn interactions and tool use.

Key Capabilities

• Bilingual Proficiency: Optimized for strong performance in both Thai (🇹🇭) and English (🇬🇧) languages.
• Extended Context Window: Features a 256K context length, enabling the processing of very long inputs and maintaining conversational coherence over extended dialogues.
• Function Calling: Supports tool use, allowing the model to interact with external functions and APIs, as demonstrated by its vLLM server deployment and OpenAI-compatible API examples.
• Instruction Following: Designed as an instruct model, it excels at following user commands and generating helpful, harmless, and honest responses.

Usage Considerations

• Deployment: Can be deployed as an OpenAI-compatible API server using vllm for efficient inference.
• Sampling Parameters: For optimal performance and to reduce repetition, it is recommended to use a low temperature and set repetition_penalty = 1.05.
• Development Status: While incorporating guardrails, the model is still under development and may occasionally produce inaccurate, biased, or objectionable content. Developers should assess these risks for their specific use cases.