DeepSeekLanguage modelSupersededOpen weights
DeepSeek V3.1
Hybrid think/no-think model with much stronger agent skills.
Overall Intelligence
#46
of 56
24.3 pts behind #1
0
races led
0
weeks at #1
64.4
score
- Released
- Aug 21, 2025
- 1.2 yr ago
- Context
- 128K
- 64K max output
- Price
- $0.56 / $1.68
- per 1M in/out · $0.84 blended
- Speed
- 45 tok/s
- 3s to first token
- Modalities
- text
- outputs text
- Weights
- Open
- 671B MoE (37B active)
The 30-second version
Hybrid think/no-think model with much stronger agent skills.
Key takeaways
- 1Hybrid think/no-think model with much stronger agent skills.
Where it ranks
Current snapshot · illustrative scores · click any race to watch it
V3.1Best on each axis
Overall
Coding
46Coding+23.9 pts47Frontend & UI+152 Elo46Backend & APIs+21.1 pts45Distributed Systems+24.4 pts44Full-Stack Apps+22.1 pts45Debugging & Repair+24.3 pts45Agentic Coding+25.4 pts46Code Review & Security+21.0 pts46Data Science & SQL+22.6 pts46Mobile Apps+21.4 pts47Terminal & DevOps+28.8 pts42Competitive Programming+610
Hard Reasoning
Math
Science
Writing & Language
Multimodal
Tool Use & Agents
Speed & Cost
Overall Intelligence over time
#1 frontierV3.1V4.1 FlashGLM-5.3
Update history
The distilled version first; open any release for the deeper breakdown.
The DeepSeek V family