Compiler

Any model. Any silicon.
Compiled just in time.

Caldrix is the next generation AI inference server.

Get startedHow it works

Why Caldrix

Compiled for your machine.

Most inference servers ship kernels hand-tuned for NVIDIA GPUs, and every other chip gets a port. Caldrix compiles the kernels, and the whole server around them, for the hardware you have and the work you give it.

Your hardware

Caldrix probes the chip, then compiles the kernels and the server for it, not an NVIDIA build ported over.

Your workload

Context length, how many requests share a step, precision and memory budget are inputs to the build.

Your traffic

As token traffic rises and falls, Caldrix re-tunes the server to match it.

How it works

Probe, compile, re-tune.

One command starts the cycle, and it keeps running: the server stays matched to the chip and to the traffic it actually gets.

  1. Probe

    Finds the chip and what it can do.

    • lane
    • ISA
    • cores

  2. Compile

    Kernels and the server, for that chip and your settings.

    • context
    • batch
    • precision

  3. Re-tune

    Rebuilds when the shape of the load changes.

    • traffic
    • KV layout
    • fusion

Why the whole server, not just the kernels: read the post.

Works on AMD · Apple · CPU — later: NVIDIA · Intel · Qualcomm

Aerix

A coding agent built around the models Caldrix runs.

Aerix works in your terminal on a local caldrix pool: each turn constrained by default, each context kept short, its prefix kept resident.

Meet Aerix

Get started

Try it on your own machine.

Install the CLI and run a model. No account needed.

Read the docs
$ curl -fsSL https://caldrix.ai/install.sh | sh
$ caldrix serve llama-8b
→ serving on http://127.0.0.1:8425