At 00:00 Beijing Time on October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017 and synchronized the official web client, App, and API to this version. On the same day, the company opened model weights as well as the Search and Fetch two API interfaces, with updates covering all three pathways: online usage, developer invocation, and self-deployment. The confirmed scope of this release includes that the new model has been launched, the weight access has been made public, and the search interface has been opened to the general public; comparisons regarding capabilities and costs are primarily based on DeepSeek's own testing and explanations.
API Interface and Model Name Remain Unchanged
DeepSeek's English update log stated that developers continue to use the model name 'deepseek-flash', with API calling methods and DeepSeek model pricing remaining unchanged. The price continuity mentioned here only applies to DeepSeek's own published API arrangements, as the announcement did not specify whether each third-party platform would adopt the same pricing. Vercel confirmed in its official change log on October 17 that the new version has entered the Vercel AI Gateway; this record indicates that another hosted calling pathway is now available, but does not publish a unified price for this channel or completely identical service conditions to DeepSeek's direct connection interface.
DeepSeek's release page provides two weight download addresses: Hugging Face and ModelScope. The Hugging Face page is under DeepSeek's official organization name, with model identification consistent with this update. The Chinese announcement refers to this section as 'model open-sourcing,' but the current materials can only confirm that weight files are available for direct download. The training data, complete training code, training process, or scope of license details are not simultaneously listed. Therefore, this article adopts the term 'open weights' and does not interpret an expansion of a model repository entry as all related materials being fully open, nor does it infer undisclosed commercial usage boundaries from the download page.
Project Development Positioning and Tool Adaptation
DeepSeek positions V4.1-Flash-1017 as a code intelligent agent and full project development, and claims that the new version adopts reinforcement learning training based on self-feedback. The official page showcases four cases including three-dimensional modeling of the Forbidden City, a web-based racing game, a recreation of Portal 2 using the Godot engine, and Valley of Ages, which presents historical evolution in a three-dimensional scenario. According to the release page, these projects were completed within the DeepSeek Harness environment with minimal prompts and materials, and shorter interactions; the sources used do not provide third-party replication records, and these four demonstrations can only serve as company showcases, unable to replace cross-task and cross-environment testing.
In terms of accompanying tools, DeepSeek recommends upgrading DeepSeek Harness to version 0.2.1-rc.1 or higher, and states that this optimization also targets platform frameworks such as Tencent CodeBuddy and WorkBuddy. The company indicates that the new version improves official benchmark scores while reducing task execution time and output length, thus referring to the results as Pareto improvements in performance-cost across multiple Harness environments. The announcement does not provide unified cost testing covering different hardware, prompt lengths, caching, concurrency levels, and third-party frameworks, therefore these efficiency descriptions should retain DeepSeek's self-reported attributes and cannot be extrapolated as identical reductions for all deployments.
Official Benchmarks and Search Interface Pricing
DeepSeek's update log lists multiple company test results: HLE no-tool score of 42.6, asterisk corresponding plain text subset score of 44.0; Terminal-Bench 4.0 at 42.3, DeepSWE v1.1 at 77.0, NL2Repo-Bench at 68.2, Automation-Bench v1.0.6 at 60.5. Official annotations also indicate that Terminal-Bench 4.0 uses the Claude Code framework, DeepSWE uses the Mini-SWE-V2 framework, and HLE asterisk score only covers the plain text subset. Different metrics come from different tasks and settings, and the sources provided do not offer independent replication of these figures.
Along with the model updates, the Search and Fetch interfaces are each priced at 5 RMB per thousand times. DeepSeek announced that when users access DeepSeek API through DeepSeek Harness, Claude Code or Codex, the server-side tools currently calling these two interfaces are currently not charged; this arrangement will end at 00:00 Beijing Time on November 1, 2026, after which billing will begin per occurrence. The company also stated that its self-developed search engine processes hundreds of millions of requests daily from around the world, indexing billions of web pages, and filtering SEO cheating, GEO manipulation and low-quality AIGC; these scales and filtering effectiveness are disclosed by DeepSeek, with no external audit results from the sources used.
Media Verification and Unverified Boundaries
QuantumBit's public information stream listed this release on October 17, 12:21:09 UTC, with the headline and short summary focusing on the new model and search API. Vercel's official records were used to confirm AI Gateway availability, while QuantumBit's public timestamp and headline were used to confirm that leading technology media had reported this event within the current window. Neither provides independent benchmark replication. The strong performance and cost-effectiveness wording in QuantumBit's headline was not taken as fact; all versions, interfaces, prices, results, and timelines mentioned in the body were verified against DeepSeek's official webpage.
As of publication, DeepSeek did not specify in this announcement whether prices across all third-party channels are consistent, nor did it fully define the code, data, training process, and licensing scope through this weight entry. Official project demonstrations, search filtering effects, task duration, and benchmark results still lack independent verification across different hardware and frameworks. Vercel availability and media report metadata cannot compensate for this gap. The next most definitive product milestone is 00:00 Beijing Time on November 1, when the fee-free window for Search and Fetch specified server-side tools will end; developers' actual billing will also depend on calling paths and usage volume.