Blog
DeepSeek releases V4.1-Flash-1017 and opens its Search and Fetch APIs
On 17 October 2026 DeepSeek released DeepSeek-V4.1-Flash-1017, an updated open-weights Flash model under the MIT License, with the same API name and price as before. It also opened its in-house Search and Fetch tools to developers at $0.75 per 1,000 calls. Benchmark and cost results are DeepSeek's own.
DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026. It is an updated version of the V4.1 Flash model the Chinese AI company launched in September. The new version is now the default Flash model on DeepSeek's website, app and API, and the weights are free to download on Hugging Face and ModelScope under the MIT License.
On the same day, DeepSeek opened its in-house web search tools to outside developers as two paid APIs, Search and Fetch.
What changed in the model
DeepSeek calls 1017 a "minor update" to V4.1 Flash. The company says it scores higher on benchmarks, finishes tasks faster and gives shorter answers than the version it replaces. DeepSeek describes this as a "Pareto improvement", meaning better results without higher cost.
According to DeepSeek's change log:
- Nothing changes for API users. The calling method and prices stay the same, and the
deepseek-flashmodel name now points to the new version automatically. - It is tuned for coding tools. DeepSeek says it optimized the model for its own DeepSeek Harness, as well as Tencent's CodeBuddy and WorkBuddy, and recommends Harness v0.2.1-rc.1 or later.
- It is trained to build whole projects. DeepSeek says it used reinforcement learning with self-feedback to improve how the model builds complete software projects from scratch. Its examples include a web racing game and a 3D model of the Forbidden City.
The model card lists a context window of 1 million tokens. It does not include a ready-made chat template. Instead, DeepSeek supplies reference code for formatting prompts and a set of Rust libraries, called deepseek-recipe, for production use.
DeepSeek's benchmark claims
DeepSeek published scores on 11 benchmarks, and the new version beats the previous V4.1 Flash on all of them. Some examples from the model card:
- DeepSWE v1.1 (software engineering): 77.0, up from 74.2
- Terminal-Bench 4.0 (command-line tasks): 42.3, up from 31.2
- Humanity's Last Exam without tools: 42.6, up from 36.8
DeepSeek's table also compares the model with Claude Opus 5, GPT-5.6 Sol, Kimi K3 and GLM-5.3. The results are mixed. DeepSeek's model leads on DeepSWE and some agent tests, while Claude Opus 5 scores higher on Terminal-Bench 4.0, ProgramBench and Humanity's Last Exam.
DeepSeek also published a cost chart. As The Decoder read it, the new model reaches its 77.0 DeepSWE score at just under $0.20 per task, compared with about $0.50 for the previous version.
All of these are DeepSeek's own results. The Decoder noted that no independent benchmarks had been published at the time of release.
Search and Fetch APIs
DeepSeek says its Search and Fetch tools run on a search engine it built itself, which it says handles hundreds of millions of requests a day. Developers can now call these tools directly:
- Search returns web results for a query.
- Fetch retrieves the content of a given web page.
Each endpoint costs $0.75 per 1,000 calls. DeepSeek API users can use Search and Fetch for free through server-side tools in DeepSeek Harness, Claude Code and Codex until billing starts on 31 October 2026 at 16:00 UTC (17:00 Lagos time).
Why it matters
The release continues a pattern for DeepSeek: frequent updates to cheaper models, published as open weights under a permissive license. Because the API name and price stay the same, developers already using deepseek-flash get the new model with no changes to their code.
The Search and Fetch APIs also matter for agent builders. Coding and research agents often need live web access, and DeepSeek is now selling that piece as well as the model. How well the model performs on real projects, outside DeepSeek's own tests, is something independent testing still needs to show.