DeepSeek-V3.1 puts thinking and non-thinking modes in one model
V3.1 merged DeepSeek's V and R lines into one model with two modes. The deepseek-chat endpoint is non-thinking and deepseek-reasoner is thinking, both on a 128K context, after 840B tokens of…
- Date
- 21 August 2025
- Who
- DeepSeek
- Confidence
- High
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): DeepSeek · Confidence: High V3.1 merged DeepSeek's V and R lines into one model with two modes. The deepseek-chat endpoint is non-thinking and deepseek-reasoner is thinking, both on a 128K context, after 840B tokens of continued long-context pretraining. DeepSeek claims V3.1-Think answers faster than R1-0528 and calls the release "our first step toward the agent era" (DeepSeek release notes). It is DeepSeek's step toward the hybrid design that Claude 3.7, Qwen3 and GPT-5 had taken, fully realised in V4. It was also the DeepSeek model that NIST evaluated as the lead DeepSeek system in its September 2025 review (B08-12).