About

Democratize the silicon.

Today, fast AI mostly means NVIDIA. We think everyone should be able to run AI well, on whatever hardware they already have.

Why we exist

The chips are ready. The software isn't.

AMD, Intel, Apple, Qualcomm and ARM all make chips fast enough to run AI well. What they lack is software tuned for them.

Tuning by hand for every chip never ends. A compiler does it once, for all of them. That means more choice, lower prices, and AI that can run offline or on a phone.

Mission

Make every chip fast at AI, and let people run models wherever they want.

Any model. Any silicon. Full speed.

How we work

Four rules you can hold us to.

Build once, run everywhere

One compiler builds the fast version for each chip it supports (AMD GPUs, Apple Silicon and CPUs today), instead of rewriting by hand for each one.

Measure, don't guess

We measure your hardware before building for it, and check every speedup with real tests.

Only say what we checked

Our tools report what they measured, and say plainly when they didn't.

Mark every unproven number

Caldrix hasn't launched yet, so every figure on this site is marked as illustrative.

Open and closed

You're never locked in.

The compiler itself is not open source. But the CLI is free, and caldrix doctor shows you exactly what it measured.

It runs on your machine, and since it speaks the OpenAI API and the Anthropic API, and builds you can copy and ship, switching away is one URL change.

Get started

Try it on your own hardware.

Read how to install the CLI and see what your machine can do.

Read the docsHow Inferno works