DeepSeek updates V4.1-Flash and opens its Search and Fetch APIs

Background

On 10 September 2026, DeepSeek released a model under the API name deepseek-flash. The company described it as the smallest model in its new architecture family, and as having native visual understanding — meaning images are handled by the model itself rather than by a separate add-on module.^[1] That release set the baseline against which the October update would be measured.

The 17 October update

On 17 October 2026, DeepSeek published an update named DeepSeek-V4.1-Flash-1017. Compared with the version it replaces, the company reported stronger benchmark results, faster task completion and shorter outputs. The model was trained with reinforcement learning from self-feedback for end-to-end project development: reinforcement learning is training in which a model's behaviour is adjusted by rewarding outputs that score well on an automatic check, and here the target was carrying a project through from start to finish rather than answering a single question.^[2]

DeepSeek described the release as a Pareto improvement in performance and cost compared with the previous version — a change in which no dimension gets worse, so better results did not come at the expense of higher cost.^[2]

Weights, APIs and billing

The weights of V4.1-Flash-1017 were published on Hugging Face and ModelScope, which means the model can be downloaded and run outside DeepSeek's own service.^[2] The same release opened DeepSeek's Search and Fetch APIs to all users at $0.75 per 1,000 calls.^[2] Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex had been free; billing for them began on 31 October 2026.^[2] DeepSeek said its in-house search engine serves hundreds of millions of requests daily worldwide.^[2]

Tooling and ecosystem

DeepSeek optimized the update for its own DeepSeek Harness and for Tencent's CodeBuddy and WorkBuddy, and recommended DeepSeek Harness v0.2.1-rc.1 or later.^[2] Alongside the release, DeepSeek showcased projects built with the update, including a 3D model of the Forbidden City and a recreation of Portal 2 in the Godot engine.^[2]

Why it matters

The release is an example of how a model family is maintained rather than replaced. The September model established the smallest member of a new architecture; the October update revised that member along three axes at once — accuracy on benchmarks, speed of task completion, and length of output — while keeping cost from rising.^[2] Shorter outputs matter in practice because they cut the amount of text a user has to read and pay for, and faster completion shortens the loop between asking for something and getting it.

Two decisions shape how the update can be used. Publishing weights on Hugging Face and ModelScope gives users an option that does not depend on DeepSeek's hosted service, while the hosted Search and Fetch APIs are priced per call, at $0.75 per 1,000 calls, which makes the cost of web access calculable in advance.^[2] Moving those calls from free to billed on 31 October 2026 also marks the point at which the search layer became a separately metered part of the product rather than something bundled into tool use.^[2] The optimization work for DeepSeek Harness and for Tencent's CodeBuddy and WorkBuddy shows that the update was tuned for particular coding environments, with a minimum recommended Harness version given so that users run a compatible client.^[2]

The training objective is worth noting on its own. Reinforcement learning from self-feedback aimed at end-to-end project development points at tasks that span many steps and files, not single prompts, and the showcased builds — a 3D model of the Forbidden City and a Portal 2 recreation in Godot — are examples of that kind of multi-step work.^[2] Behind them sits a search engine that DeepSeek says handles hundreds of millions of requests a day, the infrastructure the Search and Fetch APIs expose.^[2] Taken together, the release covers the model, the weights, the retrieval layer and the coding tools around it, which is what a user needs to assemble a working setup from one announcement.