DeepSeek, the Chinese AI startup, has released V3.1-Terminus, a fully open source large language model designed to improve coding and search tasks while eliminating language mixing errors.
Chinese AI startup DeepSeek, spun off from Hong Kong’s High Flyer Capital Management, has unveiled DeepSeek-V3.1-Terminus, an upgraded version of its V3.1 large language model. The update focuses on enhancing agentic tool use and reducing language mixing errors where Chinese words occasionally appeared in English responses.
The model is fully open source under a permissive MIT License, enabling global developers to download, modify, host, and commercialise it. It is available via Hugging Face, DeepSeek iOS/Android apps, DeepSeek API, and integrated into third-party tools such as AnyCoder and NovitaLabs serverless API.
Key improvements include enhanced Code Agent and Search Agent frameworks, stronger performance on coding and search-based tasks, and more consistent language output. In benchmark comparisons, Terminus shows clear gains in SimpleQA (96.8 vs. 93.4), BrowseComp (38.5 vs. 30.0), SWE Verified (68.4 vs. 66.0), SWE-bench Multilingual (57.8 vs. 54.5), and Terminal-bench (36.7 vs. 31.3), although results on some pure reasoning tasks, like Codeforces, were slightly lower.
DeepSeek-V3.1-Terminus operates in two modes: deepseek-chat (non-thinking, function calling, JSON output, up to 8,000 tokens) and deepseek-reasoner (thinking mode, deeper reasoning, up to 64,000 tokens). The model supports 128,000 tokens of context per interaction, allowing 300–400 pages of text in a single input/output exchange.
On pricing, the API charges $0.07 per million input tokens (cache hit), $0.56 per million input tokens (cache miss), and $1.68 per million output tokens. Developers can self-host the model using provided demo code, with minor technical fixes, such as the self_attn.o_proj parameter, planned for future releases.
As VentureBeat noted,
“The 685-billion-parameter model matched or exceeded performance benchmarks of proprietary U.S.-based systems while remaining fully open source under a permissive and enterprise-friendly MIT License, which allows for commercial usage.”
The release underscores DeepSeek’s commitment to open source AI, offering a commercially viable alternative to closed systems while encouraging community-driven experimentation, tool integration, and global adoption.
