ZML's free inference server runs your models on any AI chip

For two years the answer to "what runs my model?" has been one word: Nvidia. ZML, a Paris startup, just released a free tool to change that. LLMD is an inference server that runs open-weight language models across Nvidia, AMD, Google TPU, Apple Metal, and Intel — the same weights, the same server, whatever silicon you can get your hands on. Turing Award winner Yann LeCun, who's on the cap table, called it a "hardware independent LLM inference engine."

What actually happened

Per TechCrunch, ZML shipped LLMD on July 8, 2026. It's a serving layer that takes open-weight models and runs them at speed on whatever accelerator is underneath — no rewrite when you switch from an Nvidia box to an AMD one to a Google TPU. Founder Steeve Morin (formerly VP of Engineering at Zenly, which Snap acquired) has raised about $20 million from firms including Kima Ventures, LocalGlobe, and Kindred Capital, plus angels who know this problem cold: LeCun, Docker/Dagger founder Solomon Hykes, and Hugging Face's Clément Delangue and Julien Chaumond.

Two honest caveats. LLMD is free for now but closed-source — unlike ZML's 2024 framework, which was open — so the tool itself isn't something you own. And it's entering a crowded lane: The Next Web notes it competes with open projects like vLLM and SGLang and commercial players like Baseten. "Free for now" is a growth tactic, not a promise.

Why it matters for your business

The chip is the last hard dependency in most AI stacks. You can abstract the model, route between providers, cache prompts — and still be stuck because your inference is welded to one vendor's hardware and its wait list. A layer that runs the same open weights on five different chips turns "which GPU can I get" from a blocker into a purchasing decision.

You don't have to adopt LLMD to take the lesson. The move is to serve your models behind an interface you control, so the accelerator underneath is swappable. When the Nvidia lead time is six months, or a cloud quietly reprices its GPU instances, or a client's contract demands their data run on hardware in a specific country, that's a config change — not a re-platform. The leverage isn't the chip. It's not being cornered when the one you were counting on gets scarce or expensive.

Sources: TechCrunch, The Next Web.