DeepSeek has released V4.1-Flash-1017, demonstrating complete projects built from brief requests. One piece of product evidence remains missing: how much existing functionality survives a user’s second request for changes. A working first version supports exploring an idea, but does not establish that the project can keep being delivered through successive revisions.Release and demonstrations
The official event date is 2026-10-17; the announcement page gives no time or timezone. This article uses an October 18, 08:59 Taipei cutoff and explicitly draws on the previous day’s release within the 48-hour window. RuntimeWire first reported a possible refresh at 00:08 US Central time on October 17, then updated its account at 10:07 with confirmation of the announcement. Reporting timestamps are not deployment timestamps.Timing record
The website, app and API now use the update, while the API name remains deepseek-flash and model rates are unchanged. Existing users receive it without changing interfaces, but the model behind that familiar name has changed.Independent reporting Version mapping MIT-licensed weights are available. The model card lists 763B parameters, however; downloadable weights do not make full deployment practical on an ordinary laptop.Weights
The model card reports DeepSWE v1.1 improving from 74.2 to 77.0 and Terminal-Bench 4.0 from 31.2 to 42.3. The former uses Mini-SWE-V2; the latter uses Claude Code. These are different evaluation setups, vendor measurements rather than independent replications, and not evidence of equivalent quality on Traditional Chinese requests.Evaluation conditions
DeepSeek attributes better project building to reinforcement learning with self-feedback, illustrating it with a Forbidden City 3D model and a browser racing game. It also claims shorter task times and outputs, and recommends Harness 0.2.1-rc.1 or later. The demonstrations establish recognizable uses, not typical user success rates or a record of subsequent revisions.Method and examples
A generated prototype needs to accommodate new requirements
I favor using this capability to compare interaction designs that are not yet settled. A playable game lets someone try the controls and discover what a written request left unspecified. The first version then helps determine what to change next. That choice follows from the demonstrations without assuming the model can replace ongoing maintenance.
Consider a hypothetical user who wants different steering after trying the racing game, while keeping its scoring rules. The agent must modify the existing project rather than generate a visually similar replacement. Smoother handling would not compensate for broken scoring; the user would still need to check the result. This distinguishes creating an artifact from continuing work on it.
I would therefore have such a product retain runnable versions, show the current changes and allow unsuitable revisions to be undone. Shorter model explanations need not hurt usability. But a completion message alone gives users little basis for understanding which existing behaviors changed. Version retention and visible changes are design recommendations derived from the release, not features claimed to have shipped.
Shorter model output does not remove tool charges
The same announcement opens Search and Fetch APIs, each priced at $0.75 per 1,000 successful requests. Server-side calls through Harness, Claude Code or Codex are temporarily free for DeepSeek API users, with billing beginning November 1 at 00:00 Taipei time. That promotion is separate from model rates and does not make every search permanently free.Billing scope
Shorter outputs may reduce generation expense, while additional research, rebuilding and human corrections still consume resources. The available evidence cannot establish net savings for a typical project. I would treat the first build as a proposal people can try. A stronger delivery promise needs evidence of how much functionality survives changed requirements in the same project and how much users must repair. Those results would say more about continued usefulness than a single demonstration.
The cover reuses an existing DeepSeek brand image, not a new interface or benchmark graphic. Downloading the official event image failed because the download environment could not resolve its hostname.
Sources: