DeepSeek releases hybrid-thinking DeepSeek-V3.1

What happened

DeepSeek released DeepSeek-V3.1 on 21 August 2025 under the MIT License. ^[1][2]

It is DeepSeek's first hybrid model. One model supports both thinking and non-thinking modes, with the API names deepseek-chat mapped to non-thinking and deepseek-reasoner to thinking. ^[1]

Before hybrid models, a developer who wanted fast replies and deep reasoning had to choose between separate models. A hybrid model puts both options behind one set of weights, and the API lets callers pick the mode by model name.

Key details

The model has 671 billion total parameters, 37 billion activated, and a 128K context in both modes. ^[2]

DeepSeek-V3.1-Base extends the V3 base with two long-context phases: 630 billion tokens at 32K, ten times more than before, and 209 billion tokens at 128K, 3.3 times more. ^[2]

In non-thinking mode the model scored 66.0 on SWE-bench Verified (agent), against 45.4 for V3-0324, and 31.3 on Terminal-bench, against 13.3. ^[2]

In thinking mode it scored 88.4 on AIME 2025, 80.1 on GPQA-Diamond, 74.8 on LiveCodeBench and 30.0 on BrowseComp. DeepSeek-R1-0528 scored 87.5, 81.0, 73.3 and 8.9 on the same tests. ^[2]

DeepSeek said its thinking mode reached answers faster than DeepSeek-R1-0528 and was stronger at multi-step search-agent tasks. ^[1]

New API pricing took effect on 5 September 2025 at 16:00 UTC, ending DeepSeek's off-peak discounts. ^[1]

Impact

For API users the release changed two things at once: the model names exposed by the API now map to the two modes, and pricing changed from the date given above. Developers who had relied on off-peak discounts had to plan for the new prices.