deepseek-one-gpu runs DeepSeek-V4.1-Flash on one 96GB GPU

On October 21, 2026, a DEV Community article introduced deepseek-one-gpu, an open-source engine released by a developer for running DeepSeek-V4.1-Flash on a single GPU. According to the article, DeepSeek-V4.1-Flash has 763B parameters, and even after quantization to Q2_K it still occupies 264.5 GB, while one RTX PRO 6000 Blackwell offers 96 GB of VRAM. The engine serves the model from that single card together with the server's RAM, exposing it through an OpenAI-compatible API.

Drawing on measurements published in the repository, the article explains what determines speed. Of the 384 experts in each layer, the 5,825 most frequently used are kept on the GPU, taking up 69.5 GiB, while the remaining experts sit in pinned memory. In the measured runs, roughly nine out of ten expert selections landed on the GPU. The write-up is aimed at people who plan to set the engine up themselves, covering the factors that govern throughput and what to check before blaming the engine for slow performance.

That is where the public information currently stops: the engine is open source, the performance figures come from the project's own repository, and no independent verification or follow-up release notes have appeared in the material reviewed.