French startup ZML has launched a free AI inference server that runs open source language models across Nvidia, AMD, Google TPU, Intel and Apple hardware, giving enterprises greater flexibility while reducing vendor lock-in.
French AI startup ZML has launched ZML/LLMD, a free AI inference server designed to run open-source language models efficiently across hardware from Nvidia, AMD, Google TPUs, Intel and Apple. By enabling the same models to run across multiple chip vendors, the software aims to reduce dependence on Nvidia’s AI ecosystem and give enterprises greater freedom in choosing AI infrastructure.
The release addresses a growing challenge as AI inference—the process of running trained models to generate responses—has become the largest compute expense for many AI deployments. Cross-chip compatibility could help organisations lower infrastructure costs, adopt more power-efficient hardware and avoid being locked into a single GPU vendor.
According to ZML, LLMD is currently available free of charge, with the company prioritising adoption over monetisation. Founder Steeve Morin said the goal is to eliminate hardware silos and create a hardware-agnostic AI software layer that maximises performance regardless of the underlying chip.
“The idea is to give people back the power to create their own system.”
Morin added: “We have reached the point where we are co-designing silicon.”
Beyond large chip vendors, the software could also make it easier to deploy hardware from emerging AI chipmakers such as Axelera, Fractile, Kalray, SiPearl and VSORA. ZML, a Paris-based startup backed by AI pioneer Yann LeCun, enters a competitive AI inference market alongside open-source projects such as vLLM and SGLang, as well as commercial rival Baseten. “I couldn’t do ZML anywhere but in Paris.”
