DeepSeek Releases V4.1-Flash-1017, a Faster Update With Open Search Tools

DeepSeek released an updated version of its V4.1-Flash model on Saturday, October 17, and says the new checkpoint is faster and produces shorter answers than the one it replaces. The model, called V4.1-Flash-1017, is available on DeepSeek's website, in its app and on its developer platform. Bloomberg reporter Saritha Rai covered the launch the same day and described the change as a modest revision rather than a new generation of models.

What DeepSeek says improved.

The company announced the update in a series of posts on X and in a technical report on Hugging Face. Its own figures put the DeepSWE v1.1 coding benchmark at 77.0, up from 74.2 for the previous V4.1-Flash release. Bloomberg's summary of the company's figures puts the Humanity's Last Exam score at 42.6, up from 39.1 for the earlier version. DeepSeek did not publish speed or output-length numbers, so its claims of faster task completion and shorter outputs rest on the company's own word for now. Bloomberg noted that the earlier V4.1-Flash was offered at a steep discount, but gave no pricing for the update.

The search tools are the bigger addition for developers.

Alongside the model, DeepSeek opened its Search and Fetch APIs to all developers, which MarkTechPost describes as the first time the company has offered them publicly. The search endpoint returns ranked web results and accepts operators such as site: and before:. The fetch endpoint returns a single page as Markdown. Each endpoint costs $0.75 per 1,000 successful requests, with a limit of 5 requests per second per account. Empty results and failed calls are not charged. MarkTechPost also reports that server-side calls made through some developer tools are free until November 1, 2026.

Why it matters.

For teams building AI agents, a model maker that also sells web search and page retrieval is a notable change. Developers have often paired a model with a separate search vendor, and a per-request price is easier to estimate than a bill that depends on token counts. DeepSeek's own comparison table also lists Anthropic's Claude Opus 5 at 74.0 on the DeepSWE test. That figure comes from DeepSeek's table, not from an independent lab, and it should be read with that in mind.

What is still unknown.

The licence terms are reported differently. A third report, from tbreak on October 18, says the weights are released under the MIT licence on Hugging Face and ModelScope. MarkTechPost says the licence and hardware requirements for this checkpoint were not disclosed. None of the coverage reviewed includes independent benchmark results, and the speed and size claims have no figures attached. Developers who want to check API responses against the new endpoint can paste sample output into the free tool at /en/json-formatter. Until independent testers publish their results, the company’s numbers are the only ones available.