DeepSeek-V3.1
One 671B model with switchable thinking and non-thinking modes, stronger tool use and a UE8M0 FP8 format; DeepSeek's stated 'first step toward the agent era'.
Merges V3 and R1 behaviors via chat template, with 840B tokens of extra long-context pretraining (630B at 32K, 209B at 128K) and agent post-training. API splits into deepseek-chat (non-thinking) and deepseek-reasoner (thinking) on one model; adds Anthropic API format.
- Date
- Thursday, 21 August 2025
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
- Price
- New pricing from 2025-09-05 16:00 UTC ended the off-peak discount
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2024 (thinking) | 93.1% MMLU-Redux 93.7, LiveCodeBench 74.8 (thinking) | company |
DeepSeek's model card says the UE8M0 FP8 scale format is for microscaling-format compatibility. Base weights appeared on Hugging Face 2025-08-19 (repo creation date).
Sources
- api-docs.deepseek.com/news/news250821
- huggingface.co/deepseek-ai/DeepSeek-V3.1
- api-docs.deepseek.com/updates
This record was checked against its sources on 6 October 2026. How we check