ZML's Steeve Morin releases free LLMD to loosen AI inference from Nvidia

Why it matters

ZML is trying to turn hardware portability into a commercial wedge: if LLMD works in production, inference buyers get more leverage over chip supply, cost and vendor lock-in.

Steeve Morin is putting ZML's hardware-portability thesis into a product developers can run today: ZML/LLMD, a free inference server for open large language models across Nvidia CUDA, AMD ROCm, Google TPU, Intel oneAPI and Apple Metal.

The July 8th alpha release gives the Paris startup a sharper go-to-market wedge than its original framework. TechCrunch reported that LLMD is free at launch, closed source, and intended to help ZML learn where usage concentrates before it charges. Morin told TechCrunch he would rather measure where revenue makes sense than slow adoption by pricing too early.

Morin has earned the right to be taken seriously on hard infrastructure. Before founding ZML, he was VP Engineering at Zenly, the Paris social maps app Snap bought in 2017 for $213.3 million, according to Snap financial documents reported by TechCrunch. Earlier, Morin co-founded Veezio, a video-analysis startup that Presse-Citron profiled in 2013 as a two-person San Francisco and Paris company, and the same profile traced his path through EPITECH, INRIA, Google and Exalead. ZML is the latest version of a recurring Morin pattern: take a research-grade systems problem, push it toward production, then make the interface plain enough for other builders to use.

The product is a compiler bet dressed as a server

LLMD's first pitch is simple: run common open model families through one serving layer instead of maintaining a stack that changes by accelerator. ZML says the alpha supports Llama, Gemma, Qwen and Mistral model families. The server ships with continuous batching, paged attention, tensor-parallel sharding, prefix caching, tool calling, Prometheus metrics, and model loading from Hugging Face, S3 and GCS.

The technical detail that matters is the execution path. ZML's public repository describes a production inference stack built with Zig, OpenXLA/MLIR and Bazel, compiled directly to Nvidia, AMD, TPU and Trainium targets. The repo had about 3.8k GitHub stars when checked, giving ZML a visible developer footprint even before LLMD moved the company from framework to server.

ZML's own positioning is explicit: the company says it wants to decouple AI workloads from proprietary hardware with one codebase and many hardware targets. That is an engineering claim and a market claim. The engineering claim is that ZML can compile and run model-serving workloads efficiently across heterogeneous chips. The market claim is that buyers will care enough about portability to adopt another inference layer.

LLMD is how Morin tests both claims in public.

ZML published performance tables for Gemma and Qwen workloads across several hardware configurations. Those are company benchmarks, not third-party measurements. They still give a sense of what ZML wants operators to evaluate: time to first token, inter-token latency, total tokens per second, image size, cold-start behavior and whether the same serving feature set travels across hardware.

The product also packages platform runtimes into separate images for CUDA, ROCm, TPU, oneAPI and Metal. That packaging decision matters for a product that wants cloud operators to try unfamiliar hardware without assembling a custom runtime stack first.

Free at launch, closed source by design

LLMD's pricing is the cleanest clue to ZML's current priority. ZML already has a public open-source framework, first released in 2024 and updated with ZML/v2 on March 24th, 2026. LLMD changes the distribution model. It is a productized server, free for now, with future pricing undisclosed.

That split is rational. Open sourcing the framework helped ZML recruit technical credibility in a market where infrastructure buyers need to inspect the machinery. Keeping LLMD closed source gives Morin room to sell a managed or enterprise-grade layer later, once ZML sees how teams actually deploy it.

The risk is also clear. Inference infrastructure buyers already have strong defaults, and many are building on projects such as vLLM and SGLang or paying vendors that package model serving as a service. TechCrunch names Baseten, Inferact and RadixArk as adjacent competitors. LLMD has to prove that hardware breadth is valuable enough to offset the adoption cost of a new serving layer.

Morin's answer is portability at the hardware boundary. He told TechCrunch the goal is to give people the power to create their own systems and gain efficiency. That is founder-friendly framing, but it is also the core commercial wedge: ZML wants enterprises and clouds to use mixed chip fleets without rewriting inference code for every vendor.

The cap table says ZML is a founder bet

ZML has raised $20 million in total, according to TechCrunch. The money came from 20VC, >commit, AALVC, Drysdale Ventures, Kima Ventures, Kindred Capital, LocalGlobe and Puzzle Ventures. TechCrunch also reported founder and technical angels on the cap table, including Yann LeCun, Solomon Hykes, Clement Delangue and Julien Chaumond.

That lineup says as much about Morin as it does about the market. Investors are backing a 20-person Paris team to build infrastructure in a layer dominated by U.S. chip vendors, cloud providers and model-serving platforms. Morin's Zenly record gives him the operational credential. LeCun's presence gives ZML technical signaling power at a moment when European AI infrastructure is trying to graduate from policy talking point to shipping product.

ZML's valuation, round timing, lead investor and customer base are not disclosed. Revenue, deployment volume and retention are also absent from the public record around LLMD. Those omissions do not weaken the product announcement, but they do define what ZML still has to prove: that developers will adopt a closed-source free server, that infrastructure teams will trust it in production, and that enough of them need cross-chip portability to support a durable business.

Nvidia is the incumbent, not the only audience

The lazy version of the story would cast ZML as an anti-Nvidia startup. Morin is not making that bet. TechCrunch reported that he is not bearish on Nvidia and says ZML has a good relationship with the chipmaker. That fits the product: LLMD supports CUDA and publishes Nvidia H100 results alongside AMD, Intel, TPU and Apple numbers.

The bigger opportunity is the long tail of accelerators that are technically interesting and operationally hard to adopt. Morin told TechCrunch that ZML could help newer AI chip companies, including European vendors such as Axelera, Fractile, Kalray, OLIX, Q.ANT, SiPearl, SpiNNcloud and VSORA. That is not evidence of partnerships. It is a map of where ZML wants an edge: if a new chip can run serious inference workloads through a known software layer, the chip company gets a better shot at evaluation, and the buyer gets less vendor-specific integration work.

That is why LLMD's launch timing makes sense. The AI market has moved from model demos to production bills. Training still attracts the splashiest capital, but inference is the recurring cost center. Every prompt, tool call, agent step and generated token creates a serving bill. A company that can shave that bill, make underused hardware practical, or let a cloud provider route workloads across cheaper chips has a real opening.

Morin's challenge is that operators do not buy philosophy. They buy lower latency, lower cost, stable deployment and fewer surprises at 3 a.m. LLMD gives ZML a public artifact to test those claims against real usage. The product is free because Morin needs the data. The server is closed source because ZML needs a business. The hardware list is broad because the startup's real customer may be any team that has started to realize that model choice is only one part of the inference bill.