DeepSeek has officially released V4.1-Flash-1017, a minor update to its V4.1 Flash model, and switched its web, app and API services to the new version. API users do not need to change anything: the model is still called through deepseek-flash and pricing is unchanged, according to the company's announcement as reported by IT Home. Weights are available on Hugging Face and ModelScope under the MIT license, the same terms as V4.1 Flash.
An entry labeled 'deepseek-v4.1-flash-1017' briefly surfaced in DeepSeek's API model list on October 17 before the official release, which Pandaily reported at the time without confirmation from the company. The launch now confirms the build. Bloomberg and The Straits Times have reported the update as faster and cheaper; Chinese-language coverage adds detail on how it was trained, what it was tuned for and the service context around it.
DeepSeek describes the update as a Pareto improvement: higher benchmark scores alongside shorter run times and output lengths, lowering the overall cost per call. On Terminal-Bench 4.0, run with the Claude Code harness, the score rose from 31.2 to 42.3. ProgramBench went from 20.3 to 29.1, DeepSWE v1.1 from 74.2 to 77.0, NL2Repo-Bench from 65.4 to 68.2 and Automation-Bench from 54.8 to 60.5, according to the Hugging Face model card.
The headline capability change is what DeepSeek calls wish-style programming: building a complete, runnable project from scratch with a short prompt and very little reference material. The company attributes this to reinforcement learning based on self-feedback, in which the model critiques its own end-to-end output during training. Demos published on DeepSeek's website include a 3D model of the Palace Museum, a browser racing game, a Portal 2 recreation in the Godot engine and a cross-era 3D scene called Valley of Ages, each built in very few interaction rounds inside DeepSeek Harness.
DeepSeek also says it specifically adapted the model for DeepSeek Harness and for Tencent's CodeBuddy and WorkBuddy tools. Harness users need version 0.2.1-rc.1 or later for full support. Alongside the model, the company opened its in-house search engine to developers as standalone Search and Fetch APIs.
The release came two days after a service disruption. On the morning of October 15, users found DeepSeek unavailable for close to five hours, according to a report published on Tencent News. Based on DeepSeek's public status page, the report counted 44 recorded incidents so far this year, ranging from brief performance degradation to partial outages, including one lasting about 12 hours in late March. DeepSeek has not linked the outage to the model update.
SEE ALSO: DeepSeek Harness v0.2.1-rc1 Arrives as an Unannounced V4.1 Flash Snapshot Briefly Surfaces · DeepSeek Adds Standalone Search and Fetch APIs · OpenCode Makes DeepSeek V4.1 Flash Free to Use, With a 200K Context Window