Luminal is a high-performance, general-purpose inference compiler that squeezes maximum speed out of AI models by compiling them into highly optimised GPU code. Developed by luminal-ai and written in Rust, it is open source under the Apache-2.0 license, has about 3,030 stars on GitHub, and is under very active development in 2026. Its tagline, “inference at the speed of light,” reflects its goal: automatically find kernel implementations that hand-tuned libraries miss.
Instead of running a model op-by-op through generic kernels, Luminal represents the computation as a graph, then searches over mathematically equivalent rewrites and kernel schedules to produce a fast execution plan for the target GPU. Models are defined or imported in a Rust-native API (with crates.io distribution), lowered through the compiler’s graph optimisations, and executed with minimal overhead. The approach generalises beyond LLMs to arbitrary tensor programs, which makes Luminal interesting both as a practical inference runtime and as research into compiler-driven performance.
Luminal’s bet is that compiler search beats hand-tuning as models and hardware diversify: rather than maintaining a library of kernels per GPU generation, the compiler explores equivalent graph transformations and scheduling choices automatically for the target device. The project is developed in the open with published documentation, and its Rust implementation makes it embeddable in Rust applications as well as usable as a standalone inference engine. It is an ambitious, fast-moving project aimed at the layer of the AI stack where frameworks like PyTorch hand off to hardware.
Luminal is free and open source under the Apache-2.0 license. There is no paid cloud tier attached to the compiler itself.
ML infrastructure engineers and performance specialists who need faster inference than standard runtimes deliver, and Rust developers building AI-powered products.