A French tech startup called ZML has released free software called LLMD designed to run AI language models on computer chips made by Nvidia, AMD, Google, Apple, and Intel, according to TechCrunch. The software aims to help large companies and cloud providers avoid relying on a single chipmaker.

ZML was founded by Steeve Morin and has backing from Yann LeCun, a computer scientist who won the Turing Award, often called the Nobel Prize of computing. The product page is at zml.ai/llmd, and the underlying code is open source on GitHub.

What the software actually does

To understand this properly, you need to know the difference between the ZML framework and LLMD the product. Think of ZML as the engine — the core technology that runs AI models. LLMD is the finished car — it takes that engine and adds the practical systems needed to actually serve AI models to users, handling things like managing multiple requests at once and scheduling processing work.

Most similar software is built using Python and a language called CUDA, which are tightly connected to Nvidia chips. ZML took a different approach, using a programming language called Zig that gives developers precise control over how the computer's memory is used and keeps the software lightweight. That's a significant choice because it lets the software work better across different chip types rather than being locked into one manufacturer's ecosystem.

The software already supports chips from multiple makers. ZML released version 2 on March 24, 2026, and by then it already worked with AMD GPUs and Amazon's custom AI chips. LLMD adds support for more options — including the graphics processors in Apple computers and Intel's Arc chips — expanding the list further.

The competitive landscape

Existing companies that sell AI serving software include Baseten, and Inferact, which commercializes an open-source tool called vLLM. Both of those open-source tools have become the default choice for companies running large language models in real-world conditions over the past two years. LLMD is positioning itself as a free alternative that works with many different chips rather than competing with the underlying open-source projects themselves.

The startup is also talking to European chipmakers — eight companies including Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. These are smaller chip companies that have struggled to compete with Nvidia because Nvidia has such a dominant software ecosystem. If LLMD works well, it could let these smaller companies plug their chips into existing systems without having to build all their own supporting software from scratch.

The bigger picture

Here is where the strategy gets interesting. Nvidia has dominated AI infrastructure for over a decade not just because of chip performance but also because of the software tools surrounding it — developers have spent years learning Nvidia's tools and writing code for them. Every attempt to make truly portable software that works across different chips is an implicit bet that this Nvidia advantage can be overcome.

Whether ZML's approach will actually deliver equal performance and reliability across all these different chip types remains an open question. TechCrunch's report covers the release itself but does not include independent testing by outside reviewers, which is the real test of whether this works in practice.

The business case for companies using this software is clearer. If LLMD performs well across different chips, it gives buyers real negotiating power — they can route AI work to whichever chip is cheaper or more available at any given moment. That flexibility could reduce both cost and energy consumption compared to being locked into a single supplier.

The one practical question worth noting: ZML is a startup offering free software, and the report does not explain how the company plans to make money from LLMD itself — whether through paid hosting, support contracts, or something else. Other companies have launched similar products with an open-source core and commercial services around it, so this is not unusual. But the absence of a clear business model is worth keeping an eye on as the product develops and the company grows.

The real proof will come from whether companies actually adopt LLMD in their production systems where performance problems cost real money. If smaller European chipmakers gain even modest traction because switching between different chips became easier, that would matter in a market where software has traditionally lagged behind the diversity of available chips.