Model Drop

DeepSeek V3.1

August 21, 2025 2 min read

DeepSeek V3.1, which DeepSeek shipped on August 21, 2025, is an MIT-licensed open-weights model that puts thinking and non-thinking modes in one model and replaced both V3 and R1 behind DeepSeek's API.

Model: DeepSeek V3.1 (deepseek-ai/DeepSeek-V3.1)
Maker: DeepSeek
Ship date: August 21, 2025
Pricing: $0.56 / $1.68 per M tokens (DeepSeek API, from September 5, 2025); free weights
Context: 128K tokens
Weights: Open (MIT)
Where you can use it:
- DeepSeek chat, DeepThink toggles modes
- API as deepseek-chat and deepseek-reasoner
- Open weights on Hugging Face
When to reach for it:
- Agentic coding and terminal tasks
- Search agents that chain many steps
- Skip thinking mode for general tool calls
How it compares:
- R1 0528-level thinking, in less time
- Edges Kimi K2 and GLM-4.5 among open
What makes it different:
- One model, thinking and non-thinking
- Trained in the UE8M0 FP8 data format
Early reactions: The Register called it less a new model than an upgrade trained from an earlier V3 checkpoint, and read its UE8M0 FP8 format as a hint of new Chinese chips.
Read more: DeepSeek-V3.1 Release, model card

The benchmarks

DeepSeek's own runs, with the coding-agent rows in non-thinking mode (its internal agent framework; Terminus 1 for Terminal-Bench) and the reasoning and search rows in thinking mode; it compared only against its own DeepSeek V3 0324 and DeepSeek R1 0528. Humanity's Last Exam is the text-only subset, and the search-agent rows use DeepSeek's internal search framework (R1 0528's search scores came from a pre-defined workflow).

DeepSeek's numbers

SWE-bench Verified (real GitHub fixes, non-thinking agent mode): 66.0%. +20.6 over DeepSeek V3 0324 (45.4%), +21.4 over DeepSeek R1 0528 (44.6%).
Terminal-Bench (command-line tasks, non-thinking): 31.3%. +18.0 over DeepSeek V3 0324 (13.3%), +25.6 over DeepSeek R1 0528 (5.7%).
BrowseComp (web research agent, thinking): 30.0%. +21.1 over DeepSeek R1 0528 (8.9%).
Humanity's Last Exam (with tools) (Python and search, thinking): 29.8%. +5.0 over DeepSeek R1 0528 (24.8%).
Humanity's Last Exam (no tools) (thinking): 15.9%. −1.8 under DeepSeek R1 0528 (17.7%).
GPQA Diamond (PhD-level science, thinking): 80.1%. −0.9 under DeepSeek R1 0528 (81.0%).
AIME 2025 (competition math, thinking): 88.4%. +0.9 over DeepSeek R1 0528 (87.5%).
Aider Polyglot (multi-language code edits, thinking): 76.3%. +4.7 over DeepSeek R1 0528 (71.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 13.5. +0.4 over DeepSeek R1 0528 (13.1), +0.8 over Kimi K2 (12.7), +0.7 over GLM-4.5 (12.8), +1.9 over gpt-oss-120b (11.6).