Luminal

Luminal builds an open-source inference compiler and cloud infrastructure for AI researchers and engineering teams running custom models. Its compiler integrates with PyTorch and turns models into optimized accelerator code, reducing the manual work of preparing models for production. Luminal Cloud offers managed serverless inference with automatic batching and scaling to zero; licensed deployments let organizations run on their own infrastructure. The platform currently invites early-access signups.

Founded in 2025, Luminal’s cofounders are CEO Joe Fioti, CTO Matthew Gunton and Jake Stevens. Its Rust compiler represents models through 15 primitive operations and searches for optimized implementations rather than relying solely on hand-written kernels. Its engineering work also automates megakernel generation: combining model operations into one GPU kernel, with fine-grained synchronization to reduce idle time between operations. This builds on a technique pioneered by Hazy Research. A 2026 partnership with Positron AI extended the compiler to Atlas accelerators, demonstrating GPU prompt processing paired with Atlas token generation.

Luminal participated in Y Combinator’s Summer 2025 batch. The company reported use in research at Yale, production workloads at venture-backed startups and several research labs. In 2025, it raised a $5.3 million seed round led by Felicis Ventures to develop its compiler and inference cloud.