French startup ZML released free tool to accelerate AI inference on any chips

French startup ZML on July 8, 2026, opened free access to ZML/LLMD — a software tool for accelerating AI model inference across a wide range of hardware accelerators. The project is supported by Yann LeCun — a Turing Award laureate and Chief AI Scientist at Meta.

What ZML/LLMD Can Do

ZML/LLMD is a software layer that enables running language models faster and cheaper regardless of chip manufacturer. Cross-platform compatibility is the key advantage here: today's inference tool market is heavily skewed toward NVIDIA. Popular libraries — TensorRT-LLM, CUDA-oriented vLLM builds — are optimized by default for the "green" manufacturer's GPUs. Teams using AMD, Intel Gaudi, AWS Trainium, and other accelerators are forced to either tolerate lower performance or spend engineering resources on their own stack adaptation.

ZML/LLMD positions itself as a single tool that works equally well on different hardware and reduces the cost of running AI products in production.

• Product: ZML/LLMD — optimization layer for LLM inference
• License: free (free to use)
• Coverage: wide range of AI chips from different manufacturers
• Goal: reduce operational costs of inference in production

Why LeCun's Name Changes Perception of the Startup

Yann LeCun is one of the three Turing Award laureates of 2018 (along with Geoffrey Hinton and Yoshua Bengio) and has been Meta's Chief AI Scientist since 2018. His work on convolutional neural networks became the foundation for modern computer vision and laid the groundwork for the current wave of deep learning.

LeCun's public support is a signal to the market and investors: ZML has at least attracted the attention of top-level researchers. For developers who haven't yet heard of the startup, such a name among supporters significantly lowers the barrier of distrust toward the new tool.

Context: Why Inference Costs More Than Training

For most AI products in mature production, inference costs — literally for model responses to millions of requests — have long exceeded training costs. According to industry estimates, the cost ratio is shifting toward 10 to 90 in favor of inference. This is why optimizing the speed and cost of running models is one of the most competitive niches in AI infrastructure in 2026.

Cross-chip compatibility of ZML/LLMD potentially gives companies leverage in negotiations with vendors and reduces risks of vendor lock-in.

What This Means

ZML/LLMD addresses a real market pain point: operational inference costs grow alongside AI product scale. A free distribution model lowers the barrier to entry and can attract a broad engineering community. How much the tool outperforms specialized solutions for specific hardware will be shown by independent benchmarks as the user base grows.

*Meta is recognized as an extremist organization and banned in Russia.