DeepSeek V3.1

DeepSeek V3.1 marks the lab's first step into the agent era, built as a single model that exposes both a fast non-thinking mode and a deliberate Think mode selectable from the chat interface. The release notes describe V3.1-Think as reaching answers in less time than the prior DeepSeek-R1-0528 reasoning model, while post-training improvements sharpen tool selection and multi-step task execution, lifting results on SWE and Terminal-Bench style benchmarks and improving complex search reasoning. Under the hood, the V3.1 base was produced through 840B tokens of continued pretraining on top of V3 to extend long-context behavior, then refined with the agent-focused post-training that gives the model its practical edge.

In deployment, the model shows up through two API endpoints, deepseek-chat for direct responses and deepseek-reasoner for the thinking variant, each offering a 128K context window, an Anthropic-compatible API format, and a Beta channel with strict function calling for reliable tool orchestration. That combination makes V3.1 a good fit for builders wiring up autonomous pipelines, retrieval-augmented agents, and coding assistants that need a controllable blend of speed and deliberation, while also serving as a drop-in conversational model for users who simply want the upgraded base experience.