DeepSeek releases the DeepSeek-V3.1-Terminus update

Background

A point update keeps the same model family and parameter counts while changing training and behaviour. The name Terminus marks this release as an update to V3.1 rather than a new generation.

What happened

On 22 September 2025 DeepSeek released DeepSeek-V3.1-Terminus, an update to DeepSeek-V3.1. V3.1 had launched on 21 August 2025 as DeepSeek's first hybrid thinking and non-thinking model. ^[3][1]

The update reduced Chinese-English mixing and stray characters, and strengthened the Code Agent and Search Agent. ^[3]

Code Agent and Search Agent refer to the model's use as an agent, meaning a system that plans steps and calls tools over several turns. A Code Agent works on writing and fixing code, and a Search Agent works on multi-step web search.

Key details

DeepSeek said Terminus outputs were more stable and reliable across benchmarks than the original V3.1. ^[3]

It was available in DeepSeek's app, on the web and through the API. ^[3]

The model has 671 billion total parameters, 37 billion activated, and a 128K context in both modes. ^[2]

Impact

The Terminus weights were open-sourced on Hugging Face. About a week later, Terminus became the base for the experimental DeepSeek-V3.2-Exp. ^[3][4]

Because the update changed behaviour rather than the basic size of the model, readers comparing it with V3.1 should look at the stated changes and the benchmark results rather than the parameter count.