DeepSeek V3.1 Terminus

DeepSeek V3.1 Terminus is a large hybrid reasoning model with 671 billion total parameters and 37 billion active parameters, engineered to support both thinking and non-thinking modes. This architecture reflects a design philosophy built on the V3.1 foundation, with explicit emphasis on agentic capabilities—particularly strengthening performance in coding agents and search agents. The model targets developers and organizations that need reliable, structured outputs for complex workflows, positioning itself as a practical tool for tasks that blend reasoning with tool-driven action.

The Terminus update emerged from DeepSeek's iterative refinement process, responding directly to user feedback about language consistency and agent reliability. Benchmark improvements across reasoning benchmarks like Humanity's Last Exam (jumping from 15.9 to 21.7) and agentic tool use tasks like BrowseComp (from 30.0 to 38.5) illustrate gains in practical capability rather than purely abstract metrics. Released under the MIT License, the model is fully open source and available to the global developer community. This accessibility, combined with demonstrated improvements in agent performance and language consistency, positions V3.1 Terminus as a forward-looking choice for teams building AI-driven applications that demand both reasoning depth and operational reliability.