DeepSeek-V3.1 merges fast and reasoning modes in one hybrid model

DeepSeek/DeepSeek-V series

DeepSeek released V3.1, a single model combining fast and reasoning modes selectable by users. The update added 840B tokens of continued pretraining for long-context and agent tasks, improved tool use, and expanded API compatibility including Anthropic format support and strict function calling.

• Hybrid inference: toggle between thinking and non-thinking modes via chat interface button.
• Extended 128K token context window; updated tokenizer and chat template.
• Improved performance on SWE-bench and Terminal-Bench; better results for complex agent reasoning.
• API includes deepseek-chat (non-thinking) and deepseek-reasoner (thinking) designations.