DeepSeek released V4.1-Flash-1017 on Oct. 17, describing it as a minor update to V4.1 Flash. The release targets coding agents that use tools; DeepSeek also published the model's 763-billion-parameter weights under an MIT license and opened its Search and Fetch web APIs.
DeepSeek reports benchmark gains and publishes model weights
DeepSeek says the update is available on its website, app and API. The company also says it completes tasks faster and produces shorter outputs; those performance claims come from DeepSeek's own release and tests.
DeepSeek says the update is intended for end-to-end project development, from making a game to 3D modeling. On DeepSWE v1.1, a benchmark built around original, long-horizon software-engineering tasks in code repositories, DeepSeek reports a score of 77.0 for V4.1-Flash-1017, up from 74.2 for V4.1 Flash. Its model card identifies Mini-SWE-V2 as the test harness. This is a company-reported result for that model-and-agent setup, not a success rate for every programming task.
The model card lists 763 billion parameters, an MIT license and local-inference instructions, but no hosted inference provider for this checkpoint. The MIT license governs use of the weights; it does not provide hosted service.
Search returns results; Fetch retrieves a page
Search returns structured web-page results for a query. Fetch retrieves one URL and returns that page's content as Markdown.
DeepSeek lists each direct endpoint at $0.75 per 1,000 successful requests. Separately, the launch note says Search and Fetch calls made through DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until Oct. 31, 2026, at 16:00 UTC, when billing begins. The release does not state a post-promotion rate for those integrations.
Image: Data-center server racks photographed by Carl Lender, via Wikimedia Commons, CC BY 2.0. The image is illustrative and is not associated with DeepSeek. Used without changes. License.
.jpg){kind=link}