DeepSeek V3.1-Terminus Update Improves Tool Use, Fixes Language Errors
Chinese AI startup DeepSeek releases V3.1-Terminus, an upgrade to its open-source model that enhances agentic coding and search tasks while fixing a bug that inserted Chinese words into English text. The update is available immediately across multiple platforms.
Chinese AI startup DeepSeek introduces DeepSeek-V3.1-Terminus, an upgraded version of its open-source large language model designed to enhance performance and reduce user-reported errors. This new update focuses heavily on strengthening the model's agentic tool use capabilities, allowing it to perform noticeably better in coding and search-based evaluations. The model is available immediately through Hugging Face, the DeepSeek mobile apps, and its API, while third-party open-source tools are quickly integrating the release.
The Terminus update specifically addresses a frustrating issue from the previous V3.1 release where the model frequently inserted Chinese words into English responses. Beyond fixing this language-mixing bug, the model maintains its identity as a cost-effective, versatile workhorse for general business tasks like writing, summarization, and customer-facing chat. This contrasts with DeepSeek's R1 model, which handles deep multi-step reasoning but operates at a higher cost and slower speed.
DeepSeek first launched the 685-billion-parameter V3.1 model in August 2025 under a permissive MIT license, directly challenging proprietary U.S.-based AI systems while remaining fully open-source for commercial use. The Terminus iteration builds upon that foundation by refining the practical, day-to-day utility of the model without sacrificing its accessibility. This continuous improvement signals China's ongoing and growing influence in the competitive frontier AI development landscape.