On-device model requests

Want an open-weight model running on a phone, a Mac, or in the browser? Open a request →

I convert it, run it on real hardware, and reply in the thread with the export and the measured numbers.

The terms

Free. No invoice, no deadline, no support commitment. I take what I can, in the order I can.

The output is public. The export goes up as a public Hugging Face repo with the recipe that produced it, and the numbers land in the issue thread. That is the trade: you get a working model, I get a public record of the work.

Open weights only. I don't extract or reverse closed models.

A "no" is an answer. If a model can't be converted, I close the issue with the reason — which op broke, on which backend, and whether the wall moves. That is usually worth as much as the export.

If you need a deadline, a defined scope, or someone on the hook afterwards, that is paid work — go through john-rocky.github.io instead.

What comes back

• A public Hugging Face repo with the converted model
• The conversion script, so you can redo it when the toolchain moves
• Measured numbers on real hardware — not projections
• The variants worth having (precision / backend), with the tradeoff stated

What I convert to

Whichever fits your target — you don't have to know which:

LiteRT conversions are published to litert-community, so they sit next to the rest of the runtime's models rather than in a corner of my own namespace. That is where the LiteRT shelf below lives.

Track record

Four shelves. Every model ships with the recipe that produced it and numbers measured on real hardware.

• LiteRT — 137 conversions shipped to litert-community
• Core AI Model Zoo — 75 models
• Core ML Model Zoo — 46 models
• ExecuTorch Model Zoo — 31 models

The conversion tooling is open too: CoreML-Models · coreai-model-zoo · hf-to-litertlm

Closed issues here are the running log: every request, and what happened to it.