DeepSeek has released DeepSeek V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model, together with paid Search and Fetch APIs that let developers pull live web results into their apps. The Chinese lab published the update on 17 October and says the model is "now live on our website, app, and API," with open weights on Hugging Face and ModelScope.
DeepSeek describes the release as faster and leaner than the September version, with "stronger benchmark performance, faster task completion, and shorter outputs." For developers, the bigger change may be the search tooling, which DeepSeek now sells as a separate service.
What is new in DeepSeek V4.1-Flash-1017
The 1017 checkpoint replaces the V4.1-Flash release from September. According to a detailed write-up by Asif Razzaq at MarkTechPost, the model has:
- 552 billion backbone parameters plus 196 billion "Engram" parameters;
- 8 billion active parameters per token during prefill and 16 billion during decoding;
- a context window of 1 million tokens, with up to 384,000 output tokens.
DeepSeek's own announcement on its API documentation site does not list these figures, so treat them as reported rather than confirmed. The announcement also does not state the licence for the new weights. Developers who plan to self-host should check the model card on Hugging Face before building on it.
DeepSeek recommends running the model through its own agent client, DeepSeek Harness, version 0.2.1-rc.1 or later. It says the model is tuned for Harness and for Tencent's CodeBuddy and WorkBuddy tools.
Benchmarks: strong on coding agents, behind on hard reasoning
The scores below come from DeepSeek's own tables, as reported by MarkTechPost. Independent tests have not yet confirmed them.
- DeepSWE v1.1: 77.0, up from 74.2 for the September model. Claude Opus 5 scored 74.0 and GPT 5.6-Sol 73.0 in the same table.
- Automation-Bench v1.0.6: 60.5, ahead of every model listed.
- Apex Bench v1.1: 72.1, again the top score in the table.
- Terminal-Bench 4.0: 42.3, behind Claude Opus 5 at 51.8.
- Humanity's Last Exam (no tools): 42.6, well behind Claude Opus 5 at 56.3.
The pattern is clear. DeepSeek has pushed hard on agentic coding and task automation, where it now leads several charts. On broad reasoning and some terminal and repository tasks, the strongest closed models still come out ahead. MarkTechPost reports that the 1017 checkpoint beat the previous V4.1-Flash on all 11 benchmarks DeepSeek published, by between 2.0 and 11.1 points.
DeepSeek also says the model finishes tasks with shorter outputs. Fewer tokens per task usually means a lower bill, which is why the company calls the update "a Pareto improvement in performance and cost."
The new Search and Fetch APIs
Alongside the model, DeepSeek opened two web tools to API customers:
- Search returns live search results that a model can use to answer questions about current events.
- Fetch retrieves the contents of a specific web page.
Both cost $0.75 per 1,000 calls, according to DeepSeek. MarkTechPost adds that only successful requests are billed, that empty results and failures are free, and that each account is limited to 5 queries per second per API. DeepSeek says the tools run on an in-house search engine that already serves "hundreds of millions of requests daily."
Free until the end of October
Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users. DeepSeek says billing starts on 31 October 2026 at 16:00 UTC. Anyone testing agent workflows has about two weeks to try the search tools at no cost.
How to try it
- Use the model in the DeepSeek web app or mobile app, where it is already live.
- For API access, call DeepSeek's endpoint. MarkTechPost reports that the model is served as
deepseek-flashand works with both OpenAI-style and Anthropic-style API formats. - Read the Search and Fetch guide in DeepSeek's API docs before switching on web tools, and watch your request volume once billing starts.
- To run it yourself, download the weights from the DeepSeek-V4.1-Flash-1017 page on Hugging Face, and check the licence and hardware needs first.
Why it matters for developers on a budget
Open-weight releases have come quickly this month. We covered Reflection AI's 501B-parameter Beam model a day earlier, and it made a similar pitch about doing more with fewer tokens. DeepSeek's version adds a cheap, built-in way to ground answers in live web data, which until now usually meant paying a separate search provider.
For small teams, including startups across Africa that watch every dollar of API spend, the combination of open weights, short outputs and a low per-call search price is the part worth testing. The benchmark lead in coding agents is DeepSeek's claim. The bill at the end of the month is something each team can measure for itself.
Would you switch an existing agent workflow to DeepSeek V4.1-Flash-1017 once search billing starts, or wait for independent benchmarks first?
