DeepSeek V3.1

DeepSeek V3.1 is a 671B-parameter mixture-of-experts model with 37B active parameters per token, released in August 2025. Its defining feature is a hybrid design: a single checkpoint can run in non-thinking mode for fast chat-style answers or thinking mode for R1-class chain-of-thought reasoning, switched via the chat template. It keeps V3's 128K context window and adds meaningfully stronger tool calling and agentic post-training, beating both V3-0324 and R1-0528 on code-agent and search-agent benchmarks while holding near-R1 reasoning quality with 20-50% fewer output tokens.

The honest picture: V3.1 is a V3 architecture with R1-style post-training rather than a new generation, so raw frontier benchmarks moved only modestly. Running the full weights still needs a serious multi-GPU cluster; most users will reach it through DeepSeek's own cheap API or community hosts. The thinking/non-thinking toggle is template-driven, so providers that don't support it may expose only one mode.