DeepSeek V3.1 Terminus (Non-reasoning) is an optimized version of the DeepSeek-V3.1 language model, serving as a high-performance Mixture-of-Experts (MoE) architecture for general-purpose tasks. It is a hybrid model designed to support both thinking (reasoning) and non-thinking modes within a single weight set. This specific non-reasoning configuration is tailored for direct, low-latency inference, bypassing extended chain-of-thought processes to provide faster responses for coding, search, and assistant workflows.
The model features 671 billion total parameters, with 37 billion parameters activated per token. The "Terminus" update specifically addresses user-reported issues regarding language consistency, significantly reducing instances where the model would mix Chinese and English or generate abnormal characters. Additionally, it provides enhanced performance for agentic workflows, particularly in tool use and multi-step tasks as measured on benchmarks like SWE-bench and Terminal-bench.
Built on an expanded long-context training process, DeepSeek V3.1 Terminus supports a context window of 128,000 tokens. The model architecture remains compatible with its predecessors, allowing for efficient deployment using the FP8 data format. Its non-reasoning mode is optimized for scenarios requiring high-quality direct output, such as structured tool calling and rapid code generation, while maintaining parity with the reasoning-heavy variants in general intelligence benchmarks.