DeepSeek released V3.1-Terminus in August 2026, a hybrid AI model combining reasoning and speed in a single system. The model cuts hallucinations by 38% and improves multi-step reasoning by 43% over earlier versions. It runs on domestic Chinese accelerators instead of relying on NVIDIA chips alone. V3.1-Terminus supports two modes: Think mode for deep reasoning chains and Fast mode for quick answers, allowing users to control inference token costs. The model handles 128,000-token context windows and uses only 37 billion active parameters out of 671 billion total through mixture-of-experts architecture, resulting in lower latency and cheaper inference. DeepSeek makes V3.1 free through its API. The model outperforms DeepSeek-R1 and earlier V3 versions across benchmarks, trailing only Google’s closed Gemini 3.1-Pro on knowledge tasks but beating it on maths and coding. For open-source models, V3.1 sets a new standard, offering frontier reasoning capability without licensing fees. The model weights are free and open, but training and fine-tuning methods aren’t fully disclosed, making it more open than closed models like GPT-5 and less open than purely academic models.