DeepSeek V3.1
DeepSeek V3.1 is a hybrid open-weight model released August 2025, combining the strengths of V3 and R1 into a single system. It expands the context window to 128K tokens and introduces dual-mode operation — switching between chain-of-thought reasoning for complex queries and direct-answer mode for faster responses. Built on a 671B-parameter mixture-of-experts architecture, V3.1 achieves significant gains in multi-step reasoning, coding, and agentic search tasks. Model weights are freely available on Hugging Face.
Visit
