DeepSeek V4.1-Flash-1017: Open Weights, Benchmarks and API

A refresh under the same name

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, about five weeks after V4.1 Flash. It went live on DeepSeek's website, app and API, and the weights were published on Hugging Face and ModelScope under the MIT license. The Hugging Face card calls it a minor update to V4.1 Flash. The Decoder reported the release the same day. [1][3][2]

API users did not have to change anything. DeepSeek's documentation says the deepseek-flash model name is now served by the 1017 model with the calling method unchanged, and that requests to the retired deepseek-v4-flash names are also routed to it and billed at the Flash price. This follows the pattern of V4-Flash-0731, an earlier dated retraining released under an existing model line. [4]

DeepSeek's numbers

DeepSeek's table shows the new model ahead of its predecessor on every benchmark listed: 42.3 against 31.2 on Terminal-Bench 4.0, 77.0 against 74.2 on DeepSWE v1.1, 42.6 against 36.8 on Humanity's Last Exam without tools, and 60.5 against 54.8 on AutomationBench. The same table includes Kimi K3, GLM 5.3, GPT-5.6 Sol and Claude Opus 5. DeepSeek's model leads all four on DeepSWE (77.0 against Opus 5's 74.0) and AutomationBench, but Opus 5 is ahead on Terminal-Bench 4.0 at 51.8 and on Humanity's Last Exam at 56.3. [3]

DeepSeek described the release as a Pareto improvement: higher scores with faster task completion and shorter outputs. It said the model was trained with reinforcement learning from self-feedback to build whole projects, from games to 3D scenes. The Decoder's reading of DeepSeek's cost chart put the top DeepSWE result at just under $0.20 per task, against about $0.50 for the predecessor, and noted that no independent benchmarks were available yet. [1][3][2]

Every score here is DeepSeek's own, and the rival rows are the comparisons it chose. The table supports a narrower claim than the coverage headlines: a much cheaper model that matches or beats several larger ones on agentic coding tasks, while still trailing Opus 5 on the hardest reasoning and terminal tests. [3][2]

Search and Fetch APIs

The same post opened DeepSeek's Search and Fetch APIs, backed by its in-house search engine, to all developers at $0.75 per 1,000 calls per endpoint. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex were free until billing begins on October 31, 2026, at 16:00 UTC. [1]