DeepSeek V4.1 Flash-1017 launches, targeting coding and agent tasks

DeepSeek made its model update, V4.1 Flash-1017, available on 17 October 2026 through the company’s website, app and API. The update’s weights have also been released on Hugging Face and ModelScope. The company says the new version completes tasks faster and produces shorter responses than the previous Flash model. In the same announcement, DeepSeek introduced Search and Fetch APIs, opening its own web search tools to developers.

What tasks does the update target?

According to DeepSeek, V4.1 Flash-1017 is intended for uses such as developing software projects end to end and having AI agents carry out tasks. In the model’s post-training phase, the company uses a self-feedback reinforcement learning method it calls RLSF. Under this approach, the model’s own confidence scores or generated critiques are treated as reward signals; DeepSeek says the method does not require human evaluators or external reward models.

The company claims this method improves end-to-end project development across tasks ranging from game programming to 3D modeling and design. DeepSeek also says the model has been tuned for development environments such as DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Version v0.2.1-rc.1 or later is recommended for Harness.

How should the benchmark results be interpreted?

In DeepSeek’s own tests, V4.1 Flash-1017 scores higher than the previous V4.1 Flash version. On the DeepSWE test, the company reports a score of 77.0% at the higher setting, with an average cost of slightly less than $0.20 per task. At the lower setting, the score is reported as above 72%, with a task cost of about $0.10. The previous version is reported to have scored 74.2% at a cost of about $0.50 per run.

These figures are based on the company’s own benchmarks; independent tests have not yet been published, it notes. The results therefore do not mean that the same speed or cost will apply to every workflow. Factors such as task type and how the model is run may matter in practice. DeepSeek also argues that the new version produces shorter outputs, improving its cost-performance balance.

Search and Fetch API pricing

The Search and Fetch endpoints made available to developers are each priced at $0.75 per 1,000 calls. According to DeepSeek, the tools use the company’s own search engine, which it says handles hundreds of millions of queries worldwide each day. Setup details are available in the API documentation.

DeepSeek API users can access these features for free through server-side tools in DeepSeek Harness, Claude Code and Codex until 16:00 UTC on 31 October 2026. This date and time mark the end of the free-use period; the source does not provide details of any different pricing that may apply afterward.

Impact on architecture, interior design and visualization workflows

Claims that the model targets 3D modeling and design tasks are worth watching for architecture firms and visualization teams, particularly for prototyping, assistant-generated code and repetitive development work. However, the sources do not establish the model’s direct compatibility with any particular 3D software, its accuracy on architectural projects or the production-readiness of its output. The company’s broader performance claims should therefore be tested against real files and tasks in the office.

The release of the model weights raises the possibility of running it locally, but the sources do not provide enough detail about the update’s hardware requirements or licensing terms. These points should be verified before making a decision, particularly for teams that do not want to send project data to external services. Those choosing to use the API should also assess per-task costs and Search and Fetch calls separately.