Luminal
by Luminal
A search-based inference compiler and serverless inference cloud that compiles models into native code for GPUs and ASICs.
Luminal is an AI inference platform built around an open-source Rust compiler. The compiler searches for the fastest equivalent way to run a model on a given chip. You can call models through its pay-as-you-go serverless cloud, or license the stack for your own cloud or on-prem hardware to raise GPU utilization and cut inference cost.
Luminal is a San Francisco company in Y Combinator's Summer 2025 batch, founded by Joe Fioti (CEO, previously designed AI accelerators at Intel), Matthew Gunton (CTO, previously built infrastructure at Amazon) and Jake Stevens. Its core is an open-source inference compiler written in Rust and dual-licensed Apache-2.0 and MIT. The compiler lowers PyTorch or Hugging Face models to a graph IR built on a small RISC-style core of primitive ops. It then uses e-graphs and equality saturation to search equivalent rewrites, fusion, tiling and scheduling choices, timing candidate kernels on the target hardware instead of relying on hand-written heuristics. It can also act as a PyTorch backend through torch.compile. The public repository supports CUDA and Metal, with ROCm and other accelerator backends on the roadmap. On top of the compiler, Luminal sells an Inference OS that load-balances work across mixed CPU, GPU and ASIC clusters, plus Luminal Cloud, a serverless API with batching and scale-to-zero. Luminal raised a $5.3M seed in November 2025. On November 12, 2026 it announced a $25M Series A led by Standard Capital, with Y Combinator, Felicis, SagaVC and Liquid 2 participating. Luminal says it processes about 90 trillion tokens a month, with a significant share on Amazon Trainium chips. The company positions itself as an alternative to Nvidia-centric serving stacks such as TensorRT-LLM and to vLLM. It has not named any customers.
An ML platform or inference infrastructure team running high-volume model serving on GPUs or non-Nvidia accelerators such as Trainium, because the compiler targets hardware utilization and per-token cost directly.
Higher throughput per accelerator from ahead-of-time compiled kernels, without hand-tuning each model and chip pair.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Free, Contact for pricing
- Target Market
- CTOs, ML Platform Engineers, Infrastructure Teams, Enterprise Developers
- Deployment
- Open-source, API-based, Self-hosted
- Founded
- 2025
- Headquarters
- San Francisco, USA
Key Features
- ✓Search-based kernel compiler
Uses e-graphs and equality saturation to search equivalent rewrites, fusion and tiling, timing each candidate on the real target hardware.
- ✓PyTorch backend
Plugs in through torch.compile with a Luminal backend, so you can try the compiler without rewriting your model code.
- ✓Small RISC-style op core
Represents models with a minimal set of primitive ops, which keeps the core small enough to read and port to new hardware.
- ✓Inference OS
Schedules and load-balances inference across mixed CPU, GPU and ASIC clusters, booting or shutting down nodes as demand changes.
- ✓Luminal Cloud serverless API
Managed endpoints with automatic batching, scale-to-zero and prepaid per-token or per-image billing, with no subscription or minimum commitment.
- ✓Licensed on-prem deployment
Lets you run the stack in your own cloud or data center with dedicated engineering support, custom kernel work and tailored SLAs.
- ✓Dynamic shapes and PyTorch validation
Treats symbolic dimensions as first-class and tests compiled ops against equivalent PyTorch implementations to catch correctness errors.
Capabilities
Use Cases
- •Cutting LLM serving cost on owned GPUs
An infrastructure team compiles its open-weight model ahead of time to push GPU utilization up and serve more tokens per H100.
- •Moving inference off Nvidia
A platform team runs production inference on Amazon Trainium or other ASICs, using the compiler in place of hand-tuned kernels for each chip.
- •Low-cost hosted model API
A product team calls DeepSeek V4.1 Flash through Luminal Cloud at per-million-token prices, prepaying credit that does not expire.
- •Compiler research and custom backends
An ML systems engineer forks the Apache-2.0 and MIT Rust codebase to add a new accelerator backend or test search strategies on kernels.
- •Mixed-cluster capacity management
An operator uses the Inference OS to rebalance live traffic across a heterogeneous fleet and shut down idle nodes as demand falls.
Ideal For
Best For
- ✓Teams serving open-weight LLMs at high volume who want to raise GPU utilization and lower cost per token
- ✓Organizations running or evaluating non-Nvidia accelerators such as Amazon Trainium or other ASICs for inference
- ✓Rust and PyTorch engineers who want an open-source compiler they can inspect, extend and use as a torch.compile backend
- ✓Developers who want cheap pay-as-you-go API access to DeepSeek V4.1 Flash or Ideogram 4.5 with no subscription
Not Ideal For
- ✗Teams that need a broad hosted model catalogue today, since the public price list shows one live text model and one image model, with GLM-5.3-Flash marked coming soon
- ✗Buyers whose procurement requires SOC 2, ISO 27001, HIPAA or named data-residency commitments, because Luminal's site names no certifications
- ✗Workloads where models change often and compile time matters, since the founder said on Hacker News that compilation takes 10 minutes to 10 hours per model
- ✗Enterprises that want references, as Luminal has not named any customers publicly
Integrations
Deployment
Market Analysis
Pros
- ✓Open-source core under permissive Apache-2.0 / MIT licences, with about 3.1k GitHub stars and 3,000+ commits
- ✓Hardware-agnostic design aimed at GPUs and ASICs, with production workloads reported on Amazon Trainium
- ✓Low published API prices with no subscription, no minimum commitment and non-expiring credit
- ✓Backed by a $25M Series A from Standard Capital, with all seed investors participating
Cons
- ✗Compilation can take 10 minutes to 10 hours per model, according to the founder on Hacker News
- ✗Hacker News commenters criticised early performance numbers as low and asked for llama.cpp baselines, and headline benchmarks are still vendor-reported
- ✗Public backends are limited to CUDA and Metal; ROCm and other accelerators are roadmap items
- ✗No named customers, no published compliance certifications, and a hosted catalogue of only one live text model and one image model
- ✗The crates.io release is out of date, so you must depend on the git repository directly
Pricing
Open-source compiler
$0
- ✓Apache-2.0 / MIT dual licence
- ✓CUDA and Metal backends
- ✓Self-built from GitHub
Luminal Cloud (pay-as-you-go)
From $0.1425 per 1M input tokens
- ✓DeepSeek V4.1 Flash: $0.1425/1M input, $0.0029/1M cached input, $0.57/1M output
- ✓Ideogram 4.5 images from $0.008 to $0.20 per image
- ✓$50 minimum top-up, credit never expires
- ✓60 requests/min and 1M tokens/min default text limits
On-Prem / licensed
Contact for pricing
- ✓Licensed cloud or on-prem deployment
- ✓Dedicated engineering support
- ✓Custom kernel optimization
- ✓Tailored SLAs
The compiler is free and open source. Luminal Cloud is prepaid pay-as-you-go with a $50 minimum top-up and no subscription, metered per token for text and per returned image for Ideogram 4.5, at Ideogram's own list prices. Higher rate limits come by emailing Luminal. Licensed on-prem pricing is not published.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 6 on domains other than luminal.com.
- 1.luminal.com — luminal.comvendor
- 2.luminal.com — pricingvendor
- 3.github.com — luminal
- 4.news.ycombinator.com — item
- 5.ycombinator.com — luminal
- 6.felicis.com — luminal announcement
- 7.letsdatascience.com — luminal raises 25m series a for inference compiler and cloud
- 8.siliconreport.com — luminal raises 25 million series a led by standard capital
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Agent Router
Open-source AI gateway, formerly Envoy AI Gateway, for LLM and MCP traffic on Envoy
Featherless.ai
Serverless inference for tens of thousands of open-weight models through one OpenAI-compatible API
Merge Gateway
Managed LLM gateway with cost-based routing, failover and per-customer budgets behind a single API
Clockwork.io
GPU fault-tolerance and fabric observability software that keeps AI training jobs running through failures