DeepSeek releases V4.1 Flash-1017 with faster runs and MIT-licensed weights
On October 17, DeepSeek updated DeepSeek V4.1 Flash to V4.1-Flash-1017, switching its web app, mobile app and API to the new model and publishing the weights under the MIT license on Hugging Face and ModelScope. DeepSeek says the update improves reasoning, coding, agentic and vision benchmarks over the previous version while cutting task execution time and output length to lower per-call cost, and claims a new performance-cost Pareto optimum across several agent harnesses, with specific tuning for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. In its own technical report, the model scores 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE v1.1; DeepSeek also says self-feedback reinforcement learning strengthened building complete projects from scratch (such as 3D modeling and browser games) for more "wish-style" coding requests, and added Search and Fetch API endpoints priced at 5 yuan per 1,000 calls.
This is a notable open-weight iteration that adds search/fetch APIs and strengthens agentic, end-to-end project generation, though its benchmark edge is currently vendor-reported.