DeepSeek releases updated open coding model and makes its web Search/Fetch tools public

DeepSeek has pushed out V4.1-Flash-1017, a fresh version of its coding model, and says it performs better on benchmarks while finishing tasks faster and producing shorter answers than the previous release. The company is framing that as a Pareto improvement: more performance, less cost, or at least less output to sift through.

The model is already live on DeepSeek’s website, app, and API. It’s also tuned for tools like DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, with the company recommending DeepSeek Harness v0.2.1-rc.1 or later for the best experience. DeepSeek also points to a DeepSWE v1.1 result comparing Pass@1 against task cost on the mini-swe-v2 harness.

And this model is not just for code completion. DeepSeek says it uses reinforcement learning with self-feedback to help with end-to-end project building, from game creation to 3D modeling. The examples it highlights are broad: a 3D model of the Forbidden City, a web racing game, a Portal 2 recreation built in Godot, and a 3D scene called Valley of Ages that moves from prehistoric times to the modern era.

The other big move is more practical: Search and Fetch APIs are now open to everyone. DeepSeek says they run on its in-house search engine, which serves hundreds of millions of requests a day worldwide, and each endpoint is priced at $0.75 per 1,000 calls. Calls through server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users, with billing set to start on Oct. 31, 2026 at 16:00 UTC. The model itself is also open-source, with releases on Hugging Face and ModelScope.