DeepSeek Releases its V3.1-Terminus Update, Featuring Improvements for Language Consistency and Agent Upgrades

The new DeepSeek-V3.1-Terminus update that has now been released is based off the December 2024-debuted DeepSeek V3 model family, and it is an upgraded version of the DeepSeek-V3.1 model that came out two months back. While retaining its original capabilities, the new version aims to address certain issues that were reported previously by users.

Here’s more about it.

DeepSeek-V3.1-Terminus Update Released

As mentioned, with the newly-released DeepSeek-V3.1-Terminus version, by addressing the issues pointed out in the user feedback, improvements have been brought with respect to language consistency and agent upgrades. As noted by DeepSeek, there were mix-ups between Chinese and English texts previously, and occasionally abnormal characters were also being produced. These have been fixed with the DeepSeek-V3.1-Terminus version. Additionally, the performance of Code Agent and Search Agent have also been improved, thereby strengthening DeepSeek’s task-specific frameworks. Also, two operational modes are offered: deepseek-chat (non-thinking mode) and deepseek-reasoner (thinking mode).

Speaking more, with these improvements, notable differences have been brought on benchmarks as the DeepSeek-V3.1-Terminus now delivers even more stable and reliable outputs. In comparison to the previous DeepSeek-V3.1 version, the new DeepSeek-V3.1-Terminus has achieved better scores in MMLU-Pro (85.0 from 84.8), GPQA-Diamond (80.7 from 80.1), Humanity’s Last Exam (21.7 from 15.9), LiveCodeBench (74.9 from 74.8), BrowseComp (38.5 from 30.0), SimpleQA (96.8 from 93.4), SWE Verified (68.4 from 66.0), SWE-bench Multilingual (57.8 from 54.5), and Terminal-bench (36.7 from 31.3).

About its availability, it has now been made available on app (Android/iOS), web, and API. To add, on Hugging Face, it is immediately available, and it is also being added to AnyCoder on Hugging Face as well as to NovitaLabs serverless API.