Models

Model library

Every Myles model: where it runs, how fast, and how to get it.

One family, one job: the decision. Every Myles model reads a state, a question and options and returns a probability for every option; none writes text. Myles-G is called over HTTP; the CPU models run in your own program. Pick the size for the hardware you have, then see Model types for how to call it.

Myles-GThe hardest decisions, on our API or one GPU you own. CPU modelsMyles-400M, 150M and 22M: servers, laptops, phones and browsers.

Complete library

Model	Runs on	Median	How you call it	Available as	Status
Myles-G 31B	One data-centre GPU	48 ms	`POST /v1/decide`	Hosted API · your hardware	Available
Myles-G 12B	One workstation GPU	31 ms	`POST /v1/decide`	Hosted API · your hardware	Available
Myles-G Edge	Laptop- and phone-class GPUs	21 ms	`POST /v1/decide`	Your hardware	By request
Myles-400M	The cheapest server you have	about 500 ms	`decide()` in Python	Your hardware · evaluation weights	Available
Myles-150M	A laptop	196 ms	`decide()` in Python	Your hardware · evaluation weights	Available
Myles-22M	A browser or a phone	27 ms	`decide()` in Python	Your hardware · evaluation weights · in the browser	Available

Median time for one decision: Myles-G on one GPU; the CPU models on one thread of a desktop processor. "Your hardware" is under a licence per deployment; see Licensing or request a self-hosted evaluation.

Myles-G

Myles-G is for the hardest decisions. Launchables trains it for decisions only, and it runs on one GPU with no customer data leaving your premises. Myles-G 12B and Myles-G 31B are available on the hosted API; every Myles-G model is available by request on your own hardware. Every Myles-G model is called the same way: POST /v1/decide with a state and typed questions (API reference). Myles-G 12B and Myles-G 31B also read images sent with the request (Images).

On the 231 public JevBench items, Myles-G 31B answers 211, 12B 203 and Edge 189. Unverified: our own measurement, of the builds our API serves for the 31B and 12B (16 October 2026; the 31B reads 210 to 212 across server restarts) and of Edge on 11 October 2026; not an official placement, and official evaluation pending. Myles-G is adapted from Google's openly licensed Gemma 4 (Apache 2.0); Gemma is a trademark of Google LLC.

CPU models

Myles-400M, 150M and 22M need no GPU and run on disconnected networks. Myles-400M is 91.8% correct on our frozen set of 1,400 held-out decisions, on one CPU thread (measured on the September 2026 Myles-400M build). Organisations can request the weights of all three for internal evaluation on Hugging Face: Myles-400M, Myles-150M and Myles-22M. We review every request. Myles-22M runs in a browser: try it or read its model card. The CPU models have no HTTP API: you load the model file in your own program and call decide(state, question, options), which returns a probability per option (example).

Lineage

Every new Myles model ships with a record of its base model, data and licences. Our standard: no Chinese-origin models or data anywhere in the lineage, and no outputs of hosted models as training data. See What is in a lineage record.