AI Today: What DeepSeek gave away, and what it kept
DeepSeek has open-sourced the software that runs its AI models. Its own notes list three things you would still need to run a real service.
DeepSeek's documentation for its new open-source release starts with what it is giving away: the software it uses to run its AI models at scale. Further down, under a plain heading, it lists three things that are not included, which you would need to run a complete service.
AI Today is a series from Fíor, a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI. Each piece explains one development in AI and shows where the claims come from.
The release is called DeepInfer, and DeepSeek published it on 20 October under the Apache 2.0 licence, which lets anyone use and change it, including commercially. It handles inference: the work of running a model that has already been trained, every time someone asks it something. That is where much of the day-to-day cost of AI sits, and Bloomberg framed the release as a tool to help cut it. DeepSeek says DeepInfer is the framework behind its own production services, and that it was originally built largely on vLLM, an existing open-source project.
For now it runs one model, DeepSeek's V4.1 Flash, on two kinds of chip: Nvidia's newest data-centre GPUs and Huawei's Ascend 950.
What stays with DeepSeek
The three missing pieces are a system for keeping the model's working memory between requests, the setup that splits the stages of answering a request across separate machines, and the traffic control that spreads requests across copies of the model. Without them, DeepInfer runs what DeepSeek calls offline inference: batches of work, rather than a live service answering users as they arrive.
To DeepSeek's credit, it says this itself, in its own documentation, and adds that most of the engine work for the second piece is already done. The more careful coverage, from The New Stack and The Next Web among others, carried that distinction through.
The performance charts are DeepSeek's own, measured on 4 Nvidia GPUs for one stage and 64 for the other. DeepSeek publishes a script so that others can reproduce them. When this piece was written, no independent results had been published.
Equal billing, nearly
The New Stack's headline pointed out that Huawei gets equal billing with Nvidia. That matters. US export rules have restricted sales of Nvidia's most advanced chips to China in recent years, so software that runs well on Huawei's own chips gives Chinese AI companies another route.
Equal is not quite equal, though. Two features, keeping results identical across different batches of work and giving exactly the same output for the same input every time, are marked as Nvidia only.
DeepSeek's documentation, mostly plain lists of features and settings, read as human when run through Fíor.
Open source used to suggest you could rebuild the whole thing yourself. Here it means you can see how DeepSeek's engine works and run it in batches. The parts that turn it into a live service, and much of the advantage that comes with them, stay with DeepSeek.
Try the browser extension at Fíor.