DeepSeek open-sources the engine that runs its own AI service.

On Tuesday DeepSeek released DeepInfer under the Apache 2.0 license. It's the software that runs DeepSeek's own production service. For now it supports only DeepSeek V4.1 Flash, and only on Nvidia's newest data-center GPUs (the Blackwell generation, SM100) or Huawei's Ascend 950 chips. It isn't a complete service on its own: it leaves out the caching, the split between processing prompts and writing answers, and load balancing across machines. The speed numbers are DeepSeek's own, and no one outside has tested them yet. What this means for you: you won't run this at home. The likely effect is that companies hosting DeepSeek models can copy DeepSeek's own methods, which could make third-party access to V4.1 Flash cheaper and faster. It came out a day after the US ordered federal civilian agencies to remove DeepSeek products. Sources: DeepInfer on GitHub, TNW.