DeepSeek V3.1

DeepSeek V3.1 - Hybrid Thinking MoE with Agent-Optimized Post-Training

DeepSeek V3.1 is the first model in the DeepSeek V3 family to unify thinking and non-thinking behavior in a single model. It is post-trained on top of DeepSeek V3.1-Base and introduces a hybrid inference architecture togglable via the chat template, along with post-training optimization specifically targeting tool use and multi-step agent tasks.

Base Model Training

DeepSeek V3.1-Base is built upon the original V3 base checkpoint through a two-phase long-context extension approach:

• Phase 1: 630B tokens extending context to 32K
• Phase 2: 209B tokens extending context to 128K

Both phases use an expanded dataset of additional long documents compared to the original V3 training.

Reasoning Workflow

V3.1 introduces a hybrid thinking mode - one model supporting both thinking and non-thinking behavior by changing the chat template prefix:

• Non-thinking mode: Prefix closes the `` tag immediately, bypassing chain-of-thought
• Thinking mode: Model reasons within `` tags before answering

DeepSeek V3.1-Think is documented as reaching comparable answer quality to DeepSeek-R1-0528 while responding more quickly. API aliases: deepseek-chat → non-thinking mode; deepseek-reasoner → thinking mode.