DeepSeek has announced DeepSeek-V4.1-Flash-1017, an update focused on coding, agent tasks and complete project development. In its release announcement, the company says the model is available on its website, mobile app and API. It also links to open-source releases on Hugging Face and ModelScope.
What the company says changed
DeepSeek reports stronger benchmark performance alongside faster task completion and shorter outputs than the previous version. These are vendor-reported results. Benchmark gains may not transfer equally to every application.
The company says the release is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. It recommends upgrading DeepSeek Harness to v0.2.1-rc.1 or later for the best experience. The announcement also describes reinforcement learning with self-feedback as part of its approach to end-to-end project development.
Project examples and a separate API launch
DeepSeek shows examples including a Forbidden City 3D model, a browser racing game, a Godot-based Portal 2 recreation and a 3D scene called Valley of Ages. These are examples selected by the developer, rather than an independent comparison.
The same announcement opens DeepSeek's Search and Fetch APIs to everyone. It lists each endpoint at $0.75 per 1,000 calls. It separately says calls through server-side tools in DeepSeek Harness, Claude Code and Codex are temporarily free for DeepSeek API users, with billing scheduled to begin on October 31, 2026 at 16:00 UTC.
What builders should check
For a team considering the update, the useful next step is a small test against its own coding and tool-use tasks. Compare successful completions, latency and total cost, rather than assuming the headline benchmark result will transfer unchanged.
Model performance and the surrounding harness
The announcement's emphasis on DeepSeek Harness, Tencent CodeBuddy and WorkBuddy is a reminder that a coding model is only part of an agent workflow. The harness is the software surrounding the model: it supplies context, presents tools and carries information between steps. DeepSeek's recommendation to use Harness v0.2.1-rc.1 or later is therefore an operational detail, not simply a version number in the release notes.
For a team comparing this release with an older model, changing both the model and the surrounding workflow at once can make results difficult to interpret. Keep a record of the model version, tool settings, starting repository and task instructions. If a result improves, that record helps distinguish a model change from a different prompt, test environment or set of available tools.
What the demonstrations do and do not show
DeepSeek describes reinforcement learning with self-feedback as part of the release's approach to complete project development. Its selected demonstrations range from a racing game to 3D scenes. Those examples show the kinds of outputs the developer wants readers to examine. They do not establish a success rate for arbitrary projects, nor do they say how much manual work a different project will need.
A useful coding evaluation should look beyond whether an initial screen renders. Give the agent a bounded repository task with a clear expected behavior. Run the existing tests, inspect the changed files and try the failure cases the task should handle. For a new project, include installation instructions and a repeatable way to start it. That makes the output easier to judge than an attractive screenshot alone.
Search and Fetch have their own billing details
The release announcement also introduces access to Search and Fetch, two endpoints powered by DeepSeek's search infrastructure. DeepSeek lists each endpoint at $0.75 per 1,000 calls. That statement should be kept separate from the temporary free access described for calls made through server-side tools in DeepSeek Harness, Claude Code and Codex by DeepSeek API users.
The announcement gives a precise end to that temporary arrangement: billing starts October 31, 2026 at 16:00 UTC. A product team should not treat the current temporary allowance as a permanent operating cost. It should record how often a real task invokes search or fetch and check the current service documentation before choosing a production budget. The endpoint price also does not describe every other charge an application might incur.
Open release does not remove integration work
DeepSeek links to the model release on both Hugging Face and ModelScope. Those links provide a separate route from using the hosted website, application or API. They do not mean that every route has the same setup, hardware needs or operating costs. A team should choose its intended deployment route before comparing implementation effort.
For an API evaluation, begin with a small set of representative tasks and keep their expected outcomes. Include one task where the correct behavior is to stop and ask for missing information rather than produce more code. Record successful completion, elapsed time, output length and the actual service usage together. Shorter output is only valuable when the task still finishes correctly.
The release's practical significance is that DeepSeek is targeting complete coding and tool-use workflows, while also expanding the services those workflows can call. The useful next step is a controlled test of that combination, not an assumption that a vendor benchmark predicts every project's result. Compare agent-development workflows or read more AI News.