DeepSeek has upgraded V4.1 Flash to a new snapshot, DeepSeek-V4.1-Flash-1017, and API users don't need to change anything to get it: the model name is still deepseek-flash, and DeepSeek's change log says the calling method and prices stay the same. The company announced the release on X at 16:35 Beijing time on October 17, saying the new version is already live on its website, app and API, and it published the weights on Hugging Face and ModelScope the same day.

DeepSeek says 1017 posts stronger benchmark results than the previous V4.1 Flash while finishing tasks faster and producing shorter outputs, which it calls a Pareto improvement in performance and cost. It says the model was optimized specifically for coding agents including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends upgrading DeepSeek Harness to v0.2.1-rc.1 or later for the best results. In its announcement thread, DeepSeek also says it used reinforcement learning with self-feedback to improve how the model builds complete projects from scratch, with examples ranging from games to 3D modeling.

DeepSeek's benchmark numbers

The figures below come from DeepSeek's Hugging Face model card and match its change log. Each pair is the new 1017 score followed by the previous V4.1 Flash score. We haven't run these tests ourselves.

  • HLE without tools: 42.6 vs 36.8 (44.0 vs 39.1 on the text-only subset)
  • Terminal-Bench 4.0: 42.3 vs 31.2. DeepSeek ran 1017 with the Claude Code harness, while the original V4.1 Flash card says its coding agent scores came from DeepSeek Harness, so this pair isn't a strict like-for-like comparison.
  • DeepSWE v1.1, run with the Mini-SWE-V2 harness: 77.0 vs 74.2
  • ProgramBench: 29.1 vs 20.3
  • NL2Repo-Bench: 68.2 vs 65.4
  • Automation-Bench v1.0.6: 60.5 vs 54.8

The model card also reports gains on Agents' Last Exam, Apex Bench and three visual agent benchmarks: Chartography, BabyVision and ZeroBench.

Price, parameters and availability

Pricing hasn't changed. On DeepSeek's pricing page, deepseek-flash costs $0.3 per million input tokens on a cache miss, $0.006 on a cache hit and $1.2 per million output tokens at peak times, and half that off-peak. Peak hours are 09:00 to 12:00 and 14:00 to 18:00 Beijing time on weekdays, excluding Chinese public holidays. The context window is still 1M tokens with up to 384K output tokens, and the concurrency limit is 2,500. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but DeepSeek says those requests are now served by 1017 and billed at the Flash price.

DeepSeek's model card describes 1017 as a minor update to V4.1 Flash and doesn't list a separate parameter count. The V4.1 Flash card describes that model as a multimodal mixture-of-experts model with 552B backbone parameters that activates 8B parameters per token during prefill and 16B during decode. The 1017 weights are released under the MIT license. The model is also on OpenRouter, which says DeepSeek is the first provider, with more on the way.

The same announcement opened DeepSeek's Search and Fetch APIs to everyone, which we've covered separately. On Saturday morning, a few hours before the announcement, the third-party tracker account @aitrackerbot had reported briefly seeing the deepseek-v4.1-flash-1017 name in DeepSeek's API model list.