DeepSeek V4.1 Flash: a faster open model for whole projects
The new version was post-trained with reinforcement learning for end-to-end development, runs faster and gives shorter answers. Weights are on Hugging Face.
DeepSeek has released DeepSeek-V4.1-Flash-1017. The company describes the update as an improvement across the board, in both performance and cost.
What’s new
Compared with the previous version, DeepSeek says the model:
- scores higher on benchmarks;
- completes tasks faster;
- gives shorter outputs.
It was trained with reinforcement learning and self-feedback to run projects from start to finish better, including game creation and 3D modeling. The company’s demos include a 3D model of the Forbidden City and a web racing game.
How to try it
The model is live on DeepSeek’s website, app and API. Open weights are on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope. For coding, the company recommends DeepSeek Harness v0.2.1-rc.1 or later.
Limitations
Benchmark results appear only in charts, with no numbers in the text. The announcement doesn’t state the license, model price or context size. Watch costs too: Search and Fetch calls through server-side tools are free only until October 31 at 16:00 UTC, then cost $0.75 per 1,000 calls per endpoint.
Prepared with AI from public sources.