Introducing CHOMATO: A lightweight harness for LFM 2.5 with superpowers.

Have you ever wondered how cool it would be if your harness could:

• Branch the KV-cache without huge buffers.
• Attach precalculated context blocks however you please.
• Give perfect structured responses in any shape you'd like.
• Run on the backend, frontend, or anywhere WebGPU works.
• Do all of that in less than 1 GB of RAM.

I definitely wanted one, and since there wasn't any — I built my own.

The GUI is mostly for diagnostics (styled after Classic Mac OS), but you can check out the live online demo here: https://3ksoft.github.io/chomato/

Code: https://github.com/3ksoft/chomato (AGPL-3.0). It also uses two libraries I've authored — they aren't on npm yet, but current versions are available on my GitHub under the MIT license.

The API is a bit unusual if you're used to traditional ones:

const result = await engine.generate(
  type({ id: "number", name: "string < 64" }),
  { checkpoint, blocks },
);

Basically, sparse (structured) mode is the engine default. If you just want plain text, you use type("string < max_length").

Hope you'll have as much fun using it as I had while building it!