DeepSeek-V4.1-Flash-1017 model weights are free to download under MIT

DeepSeek has released DeepSeek-V4.1-Flash-1017, a minor update to DeepSeek-V4.1-Flash, and the pitch is straightforward: better results, less drag. The company says it improves reasoning, coding, agentic tasks, and visual benchmarks, while also cutting task execution time and output length.

That combination matters because it pushes on both sides of the usual tradeoff. The model is meant to finish work faster and at lower cost, and DeepSeek says that helps move the performance–cost Pareto frontier across multiple agent harnesses. In other words, it is not just about higher scores; it is about getting those scores with less work on the way there.

The benchmark sheet backs up that claim in a very specific way. DeepSeek-V4.1-Flash-1017 is ahead of DeepSeek-V4.1-Flash on every benchmark listed in the report, including HLE without tools, Terminal-Bench, DeepSWE v1.1, ProgramBench, NL2Repo-Bench, Automation-Bench v1.0.6, Agents’ Last Exam, Apex Bench v1.1, Chartography, BabyVision, and ZeroBench-main. Some of those are text-only, some use tools, and the report notes that Terminal-Bench 4.0 used the Claude Code harness while DeepSWE v1.1 used Mini-SWE-V2.

DeepSeek is also making the practical side unusually explicit. The release does not include a Jinja chat template, but it does ship a self-contained Python reference implementation in encoding.py, plus deepseek-recipe, a set of Rust libraries with Python bindings for prompt formatting and parsing. For people who want to run it locally, the inference folder covers weight conversion and inference instructions. The recommended sampling setup is temperature 1.0, top_p 1.0, a 1M-token context window, and max_tokens of at least 256K.

And yes, the weights and the repository are both under the MIT License. That makes this less like a teaser and more like a proper open release, which is still the cleanest way to matter in a field full of glossy demos and closed doors.