DeepSeek releases V4.1-Flash-1017 and opens Search and Fetch APIs

DeepSeek describes V4.1-Flash-1017 as stronger on benchmarks, faster at completing tasks and shorter in its outputs than the previous version.[1] The company calls the checkpoint a Pareto improvement in performance and cost.[1] In its benchmark reporting, DeepSeek says the new version beat the previous V4.1-Flash on all 11 benchmarks it reported, with gains of 2.0 to 11.1 points.[2] MarkTechPost lists a score of 77.0 for V4.1-Flash-1017 on DeepSWE v1.1.[2] These are reported comparisons, not evidence here of performance on a particular team’s workload.

DeepSeek says the checkpoint uses reinforcement learning with self-feedback to improve end-to-end project development, covering tasks from game creation to 3D modeling.[1] It says the model is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.[1] For teams using Harness, DeepSeek recommends version 0.2.1-rc.1 or later.[1] The announcement does not provide task-specific results for these tools, so teams considering a switch may want to test their own development workflows before relying on the model’s reported gains.

DeepSeek says its Search and Fetch APIs are now open to everyone, and that its in-house search engine serves hundreds of millions of requests daily worldwide.[1] Each API costs $0.75 per 1,000 successful requests.[3] Billing begins October 31, 2026, at 16:00 UTC.[1] Search supports requests for between one and 50 results.[4] These terms give teams a basis for estimating request costs, but the cited information does not specify the meaning of every value listed in the API documentation.

DeepSeek lists V4.1-Flash-1017 with a release date of October 17, 2026, and says it is live on its website, app and API.[1] MarkTechPost reports 552 billion backbone parameters and 196 billion Engram parameters for the model.[2] It also reports that the model activates 8 billion parameters per token during prefill and 16 billion during decoding.[2] DeepSeek’s documentation lists the checkpoint as a model version.[3] These specifications may help technical teams identify the release, but they do not by themselves establish how it will perform in a particular deployment.