DeepSeek drops V4.1-Flash with RL-driven coding and open weights

DeepSeek-V4.1-Flash-1017 is an open-source coding model built to take development projects from zero to done. It is now live across DeepSeek's website, mobile app, and API, with model weights freely available on Hugging Face and ModelScope.

Its main hook is end-to-end development. Instead of just suggesting snippets or autocompleting functions, the model is tuned to scaffold and generate full projects. Official demos include 3D modeling of the Forbidden City, a browser-based racing game, a playable Godot remake of Portal 2, and an interactive 3D timeline scene called Valley of Ages.

Under the hood and integrations

The upgrade relies on reinforcement learning with self-feedback. DeepSeek claims this training loop cuts down output tokens and finishes coding tasks faster than the previous release, delivering stronger benchmark numbers without bloating inference costs.

On the tooling side, the model comes tuned out of the box for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. For those running DeepSeek Harness, upgrading to v0.2.1-rc.1 or later is recommended to get the intended prompt handling and tool-use performance.

Alongside the model, DeepSeek opened up its internal Search and Fetch APIs to third parties. Powered by its proprietary search engine, both endpoints cost $0.75 per 1,000 calls. Through October 31, 2026, at 16:00 UTC, calls made via server-side tools in DeepSeek Harness, Claude Code, and Codex are completely free for API users.

Why it matters

If you build games, interactive web prototypes, or independent coding tools, this release is worth testing immediately. Models that generate full Godot projects or 3D scenes without heavy human glue code are rare, and finding one with open weights is rarer.

By releasing the weights on Hugging Face while packaging cheap, dedicated search endpoints, DeepSeek is handing developers the complete toolkit to run self-sufficient coding agents without vendor lock-in.