Luminal

Luminal is a San Francisco-based AI infrastructure company building a compiler and runtime stack that makes machine learning models run faster on GPUs, CPUs and ASICs. Rather than interpreting models dynamically at runtime as conventional inference engines do, Luminal lowers a model (from PyTorch or Hugging Face) into a minimal graph intermediate representation — a pure dataflow graph — then applies hardware-aware fusion, tiling, memory-planning and scheduling passes before emitting GPU kernels or ASIC instructions ahead of time. The company describes this compiler-first design as eliminating runtime overhead and claims 2-3x higher throughput than existing inference engines on standard benchmarks.

The open-source core, published as the Rust crate luminal-ai/luminal, is described as a high-performance general-purpose inference compiler. Its architecture reduces all computation to 15 primitive operations (unary, binary, reduction and data-movement ops), models dynamic shapes as symbolic dimensions, and integrates with PyTorch as a torch.compile backend. Instead of hand-written heuristics or destructive rewrite rules of the kind used by XLA, torch.compile and TVM, Luminal treats optimization as a search problem: it generates large numbers of candidate graphs and kernels via rewrite rules and selects the fastest by measured runtime, an approach the company says can automatically rediscover complex optimizations such as Flash Attention. The repository documents native PyTorch support, integration of kernel libraries such as FlashInfer and cuBLASLt, low-precision dtypes (mxfp4, nvfp4, fp8), and multi-device parallelism topologies searched ahead of time.

On top of the compiler, Luminal markets a 'Hyperscale Inference OS' that schedules and load-balances inference workloads across heterogeneous clusters of CPUs, GPUs and ASICs, monitoring utilization, redistributing work in real time and booting or shutting down nodes as demand fluctuates.

Founding story

Co-founder Joe Fioti was working on chip design at Intel roughly three years before the November 2025 funding announcement when he concluded that the binding constraint on hardware adoption was software, not silicon: 'You can make the best hardware on earth, but if it's hard for developers to use, they're just not going to use it.' He founded Luminal in 2025 with Jake Stevens, previously at Apple, where he worked on iPhone imaging and had earlier founded a company with an exit and served as head of growth at a startup he grew to roughly $5M ARR, and Matthew Gunton, previously an engineer at Amazon working on systems that automatically detected and remediated issues in the global fulfillment and inventory network. Fioti's Intel work included AI accelerators and CPU microcode.

Business model

Luminal sells compute and inference capacity, positioning itself similarly to neo-cloud providers such as CoreWeave and Lambda Labs but differentiating through compiler-level optimization that extracts more performance from the infrastructure it operates. Its site offers two commercial paths: Luminal Cloud, a managed serverless inference service billed on usage with scale-to-zero and automatic batching; and on-prem or licensed cloud deployment with dedicated engineering support, custom kernel optimization and negotiated SLAs. The core compiler is distributed as an open-source project on GitHub under MIT/Apache licensing.

Usage-based pricing for managed serverless inference on Luminal Cloud ('pay only for what you use', scale to zero), and licensed on-prem or dedicated cloud deployments bundled with engineering support, custom kernel optimization and SLAs. The underlying compiler is open source.

Traction

The GitHub repository shows roughly 2.9k stars, 222 forks and 2,878 commits. The YC launch post states Luminal already powers research at Yale, production workloads at VC-backed startups and several research labs, and includes a runnable Llama 3 8B CUDA example. Vendor benchmarks claim Q8 Llama 3 8B at about 80% of theoretical maximum performance on an H100 and 36k tokens/sec on GPT-OSS 120B across 8xH100 SXM versus 28k for TensorRT-LLM and 26k for vLLM. Team size is listed at seven with three open engineering roles.

Latest developments

On November 17, 2025, Luminal announced a $5.3 million seed round led by Felicis Ventures with angel investment from Paul Graham, Guillermo Rauch and Ben Porterfield. A third-party aggregator lists total funding of $5.5M across two rounds, adding a $500K seed dated September 2025 with participation from A Capital, Album VC, Craft Ventures, Crosslink Capital, Gradient Ventures, Hack VC, Innovation Endeavors, Liquid 2 Ventures, Magnetic Ventures, Pareto Holdings, Pelion Venture Partners, TSVC Capital, UpHonest Capital, Y Combinator and several angels. The company is actively hiring compiler and cloud inference engineers in San Francisco.