On the evening of September 22, 2025, DeepSeek confirmed in official documentation and changelogs that its online models have been uniformly upgraded to DeepSeek-V3.1-Terminus. This update covers both the deepseek-chat (non-thinking mode) and deepseek-reasoner (thinking mode) model lines, with both versions now labeled Terminus. The context length remains at 128KB, and users can now experience it simultaneously in both the API and client. According to the update notes, V3.1-Terminus builds on the core capabilities of the V3.1 upgrade released on August 21st, while further optimizing language consistency to reduce issues with Chinese and English mixing and unusual symbols. Regarding agent capabilities, the Code Agent and Search Agent have been enhanced. Output specifications have also been clarified: non-thinking mode defaults to 4KB, with a maximum of 8KB; thinking mode defaults to 32KB, with a maximum of 64KB. Pricing and feature support remain largely the same as before.
FAQ
Q: When was DeepSeek-V3.1-Terminus released?
A: September 22, 2025. The official update log and news release both indicate the release date.
Q: What is the relationship between V3.1-Terminus and the previous V3.1?
A: This is a version identification update for V3.1, continuing the hybrid inference architecture and functionality of V3.1. It is not a completely new series.
Q: What is the difference between thinking mode and non-thinking mode?
A: Non-thinking mode corresponds to deepseek-chat, which has faster output but shorter length. Thinking mode corresponds to deepseek-reasoner, which has stronger inference capabilities and supports longer output.
Q: What changes does this update bring?
A: Mainly language consistency improvements, as well as optimizations of Code Agent and Search Agent.
Q: What are the context and output length specifications?
A: The context is unified at 128KB; the default for non-thinking mode is 4KB and the maximum is 8KB, and the default for thinking mode is 32KB and the maximum is 64KB.
