DeepSeek-V3.1 adds hybrid thinking mode and improves tool calling

DeepSeek has released DeepSeek-V3.1, a hybrid model that supports both thinking and non-thinking modes via chat template changes. The update brings improvements in tool usage performance and faster response times compared to previous versions.

• Hybrid thinking mode allows switching between thinking and non-thinking modes by changing the chat template.
• Post-training optimization significantly improves performance in tool usage and agent tasks.
• DeepSeek-V3.1-Think achieves comparable answer quality to DeepSeek-R1-0528 while responding more quickly.
• The model is trained using UE8M0 FP8 scale data format for compatibility with microscaling formats.

The release provides instructions for using the model with libraries like Transformers, inference providers such as HuggingChat, and local apps including vLLM and SGLang.