AI
French startup ZML launches free inference engine for multi-chip AI deployment
ZML has released an open-standards inference server that lets large language models run efficiently across Nvidia, AMD, Google, Apple, Intel and other chips—breaking vendor lock-in and challenging the AI inference market.
TechCrunch — AI
Paris-based startup ZML has introduced ZML/LLMD, a newly launched inference server that enables large language models to operate across a wide range of processors—from industry incumbents like Nvidia and AMD to Google's TPUs, Apple Metal accelerators, and Intel Arc chips. The platform launches as a free product, positioning itself against mounting concerns about AI deployment costs and vendor dependency.
The case for hardware flexibility#
ZML founder Steeve Morin articulates an increasingly urgent challenge in AI infrastructure: inference workloads have become a critical bottleneck, yet deployment remains fragmented across incompatible stacks that lock enterprises into single suppliers. His vision for ZML/LLMD centers on dismantling these barriers, allowing organizations to mix and match hardware based on cost, power consumption, or performance trade-offs rather than being constrained by software compatibility.
The software accomplishes this by optimizing inference—the computational process of running a prompt through a trained model—to achieve maximum throughput across heterogeneous architectures. For enterprises scaling AI, this flexibility translates to material efficiency gains and reduced lock-in pressure.
Backing from AI leadership#
ZML has drawn backing from prominent figures in machine learning and startup ecosystems. Turing Award winner Yann LeCun has publicly endorsed the company, while its cap table includes co-founders of Docker and Dagger (Solomon Hykes), leaders from Hugging Face (Clément Delangue and Julien Chaumond), and LeCun himself through his venture vehicle. Morin, who previously served as VP of engineering at Zenly before its acquisition by Snapchat for nine figures in 2017, raised $20 million from venture backers including 20VC, >commit, Kima Ventures, and others.
A competitive market with broad potential#
The inference optimization space has become densely competitive. Rivals include Baseten, valued at $13 billion; Inferact, which commercializes the open-source vLLM project; and RadixArk, the commercial entity behind SGLang.
