Most labs open the model and keep the engine. On October 20 DeepSeek did the other half. DeepInfer is the software that actually serves its models, and it is now on a public license.
What you can run
The first drop is offline inference for DeepSeek V4.1 Flash. The chips named in the release are Nvidia’s SM100 GPUs and Huawei’s Ascend 950. Apache 2.0 is the license, which is the one infrastructure teams know how to take to a lawyer.
DeepSeek’s notes describe the unglamorous pieces that make a large mixture-of-experts model affordable: splitting attention and experts across machines, balancing the load, and a speculative decoder that drafts several tokens and then checks them. A public Q&A even mentions production scale, including decode at large expert-parallel widths and a very large cache. Those numbers are the company’s, so treat them as a claim until you measure your own cluster.
Why a serving repo matters
If you already run DeepSeek weights through another engine, this does not force a migration. It does mean you can compare against the stack the lab uses, instead of guessing which flag they forgot to mention in a blog post.
It also draws a brighter line through an awkward week. The same lab whose weights governments are arguing about just published the machine that runs them. Policy and GitHub are now moving on different clocks.