AI Running on AMD Chips Just Got Faster and Cheaper

vLLM is a free, open-source program that companies use to actually run AI models and answer your questions quickly. Think of it as a restaurant kitchen: the AI model is the recipe, and vLLM is the kitchen equipment that decides how many orders get cooked per hour. This update, version 0.30.1, is mostly about making that kitchen work with a new kind of stove.

Specifically, the developers added support for AMD's MI355 data-center chip and something called NVFP4. NVFP4 is a way of storing AI's internal numbers in a more compressed form — like saving a photo as a smaller file. The picture looks nearly identical, but it takes up far less space and moves around much faster. Less memory used per question means companies can serve more users on the same hardware, which usually pushes prices down.

The other half of the change is plumbing developers rarely mention but rely on constantly: automated testing. The project added 'mirrors' for its test systems, meaning the software is checked on AMD hardware too, not just Nvidia's. Nvidia dominates AI chips today, and having a real alternative matters. When two companies compete, buyers win — and buyers include every business renting out AI to you.

Here's the honest catch. This is a release candidate, meaning a test version, not something most people will notice. It doesn't improve the quality of AI answers at all, and it does nothing for your phone or laptop today. Its effect shows up months later, quietly, as AI services get slightly cheaper to operate and slightly less dependent on a single chip maker.