ZML Releases Free Inference Server That Runs on Almost Any Chip
Paris-based ZML released LLMD this week, an LLM inference server designed to run open source models across Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc hardware. The product is free. Monetization plans are unspecified, pending usage data.
The Chip Problem They're Solving
Most inference tooling is effectively Nvidia tooling with token support for alternatives bolted on. ZML is positioning LLMD differently: chip-agnostic from the start, with the company claiming it has reached the point of co-designing silicon with chip partners.
The European chip partners named include Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud, and VSORA. Whether those partnerships produce meaningful volume is an open question.
Who's Behind It
ZML has 20 people and raised $20 million from 20VC, >commit, AALVC, Drysdale Ventures, Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures. Notable angels include Hugging Face co-founders Clément Delangue and Julien Chaumond, Docker and Dagger founder Solomon Hykes, and Turing Award winner Yann LeCun, currently at AMI Labs.
That is a credible investor list for a 20-person team. The Hugging Face co-founders in particular have specific opinions about the inference layer.
ZML's first public work was an inference-focused ML framework released in 2024 and updated in March 2026. LLMD is its first product aimed at end users.
The Inference Market Context
The inference infrastructure space has consolidated significant capital. Baseten recently reached a $13 billion valuation. Inferact was founded by the team behind vLLM. RadixArk is the commercial entity backing SGLang.
Each is chasing a version of the same problem: running inference at scale, efficiently, profitably. ZML's angle is the chip-agnostic layer, which would matter more if non-Nvidia hardware were more widely deployed in production. Right now it mostly isn't.
Free Until It Isn't
Releasing a product free with monetization TBD is a reasonable way to gather usage data before pricing. It's also how you compete against established tooling with a 20-person team.
The real question is whether LLMD's multi-chip support drives enough adoption to build a business around. The chip landscape could diversify quickly, or Nvidia could hold its position for another five years. ZML is betting on the former.
Source: Techcrunch