Skip to content

Models

Every option, weighed in one pass.

Give Myles a situation, a question and the possible answers. It returns a probability for each, and never writes text.

Start with where it will run.

Six sizes, from browser to GPU.

Myles-G

Myles-G 31B

Our strongest model, for the decisions that need the most judgement.

Runs on
One data-centre GPU
Median decision
48 ms
Status
Available

Myles-G 12B

The balance of accuracy and cost for decisions at volume, from text or images.

Runs on
One workstation GPU
Median decision
31 ms
Status
Available

Myles-G Edge

A compact Myles-G that decides close to your data, with no network round trip.

Runs on
Laptop- and phone-class GPUs
Median decision
21 ms
Status
By request

Myles

Myles-400M

Our most accurate CPU model, for long messages and documents on ordinary servers.

Runs on
The cheapest server you have
Median decision
About 500 ms
Status
Available

Myles-150M

A middle size for quick decisions on everyday hardware, with no GPU needed.

Runs on
A laptop
Median decision
196 ms
Status
Available

Myles-22M

A tiny model for short messages, small enough to run in a browser or on a phone.

Runs on
A browser or a phone
Median decision
27 ms
Status
Available

Median time for one decision: Myles-G on one GPU, the others on one CPU thread of a desktop processor.

Keep what is certain. Hand over the rest.

  1. Decide

    Every option gets a probability in one pass.

  2. Certify

    A threshold fixes the error rate on held-out data.

  3. Hand over

    Anything below it goes to a person.

  • 5% target2.7% error

    73% kept

  • 2% target1.4% error

    40% kept

  • 1% target0.5% error

    16% kept

Decisions Myles keeps. The rest go to a person.

Myles-400M on unseen test data, September 2026.

Myles-G 12B also reads images.

  • 89% correct on 934 held-out images at the default image size
  • About a fifth of a second median decision with one image

Internal measurement on 934 held-out images, 16 Oct 2026; not an official benchmark placement.

Measured, then claimed.

Test sets are frozen before any comparison and never trained on. Each threshold is set on one held-out set and checked on another that the model has never seen. All figures are our own measurements.

  • 91.8% correct on 1,400 held-out decisions
  • £4 per million decisions on one CPU thread
  • 74% on multi-rule problems from unseen domains, where chance is 16%
  • 0 cloud round trips

Myles-400M on our frozen evaluation set, September 2026. Method and full results

Three ways to get Myles.

  • Hosted API

    Call Myles-G on our servers with prepaid credit.

    Get an API key

  • Gated access

    Run the small CPU models on your own hardware, once we approve your request.

    Request access in the console

  • Self-hosted licence

    Run any Myles in production, with one licence per deployment.

    Talk to us