DeepSeek releases DeepSeek-V3.1 with hybrid reasoning mode

DeepSeek released DeepSeek-V3.1, an update that folded its separate thinking and non-thinking model lines into a single set of open weights, selectable through the API: the deepseek-chat endpoint ran the non-thinking mode and deepseek-reasoner ran the thinking mode, both sharing one underlying model rather than requiring two. Both modes supported a 128,000-token context window. DeepSeek said the thinking mode reasoned faster than its prior R1-0528 reasoning model, and reported gains on SWE-bench and Terminal-bench alongside improved multi-step tool use, positioning the release around agentic coding tasks rather than pure question-answering benchmarks.

Base and instruction-tuned weights were published openly on Hugging Face, continuing the pattern set by DeepSeek’s earlier V3 and R1 releases of shipping frontier-competitive open weights on a rolling basis rather than a single annual flagship. Alongside the model release, DeepSeek announced new API pricing taking effect two weeks later that removed the off-peak discount it had previously offered.

DeepSeek-V3.1 arrived amid a broader 2025 trend of hybrid reasoning becoming a default architecture choice: rather than maintaining separate fast and slow-thinking model families, labs including DeepSeek, Alibaba’s Qwen team and eventually Google converged on single models that could toggle reasoning depth per request, trading a small amount of engineering complexity for one model to maintain instead of two. For DeepSeek specifically, the release kept the company’s open weights close to the frontier established by closed, better-resourced labs, sustaining the pressure its January 2025 debut had put on assumptions about how much capital a competitive model required.