DeepSeek releases V3.1: Introducing hybrid reasoning architecture, one-click switching between thinking/non-thinking dual modes

AIStarMap Editorial Department

On August 21, 2025, DeepSeek released V3.1 under the MIT protocol. It adopts a thinking/non-thinking dual-mode hybrid architecture and additionally trains over 800B tokens based on V3. It has greatly improved compared to the previous generation on SWE-bench, Terminal-bench and other benchmarks, and enhanced Agent capabilities.

On August 21, 2025, DeepSeek released under the MIT license, bringing a hybrid inference architecture.

DeepSeek-V3.1 hybrid inference architecture dual-mode release overview

Picture: DeepSeek official website dialogue interface. V3.1 trains an additional 800B tokens on the basis of V3, adopts a thinking/non-thinking dual-mode hybrid architecture, and improves by more than 40% on benchmarks such as SWE-bench and Terminal-bench compared to the previous generation. It was updated to V3.1-Terminus on September 22.

Version overview

Project Content
Copyright: Content sourced from DeepSeek official . This platform has compiled and organized this content for informational purposes and learning exchange only. If there are any copyright concerns, please contact us for resolution.

Reviews

  • Loading reviews...