Model card
Myles-22M
A Launchables model · Version 2 · October 2026
Overview
Myles-22M is a tiny decision model built by Launchables. Give it a message, a question and a set of options, and it returns a probability for every option in a single pass. It doesn't write text; it decides.
With 22.7 million parameters in a 23 MB model file (about 34 MB including the runtime that runs it), it runs entirely in your browser, on laptops and phones, and answers in tens of milliseconds. Nothing you type is sent anywhere by this site. It's named after Myles.
What it's good at
- Spotting scams and phishing
- Judging urgency and routing messages to the right team
- Detecting personal details in text
- Reading tone and intent: happy or upset, cancel or upgrade, bug report or refund
- Answering your own yes/no and multiple-choice questions about a short message
Evaluation
Measured on data held out from training.
| Test | Result | Random guess |
|---|---|---|
| 1,400 frozen held-out decisions (average of four sets) | 76.3% | 38.4% |
| Demo questions with unseen phrasing (1,800) | 97.6% | 38.1% |
| General decision benchmark, test (6,000) | 62.5% | 35.8% |
| General decision benchmark, out of distribution (6,000) | 64.6% | 37.4% |
| Yes/no answers that change when the options are listed in the other order (579) | 1.2% | — |
Full benchmark results, including cost, speed and accuracy against hosted AI models, calibration and an independent decision-model benchmark, are on the Myles-22M page on Hugging Face .
Limitations
Myles-22M is experimental and can make mistakes, sometimes confidently. It's strongest on short messages and weakest on long documents or questions that need several steps of reasoning. Don't rely on it for anything important.
How it works
A compact transformer encoder reads the question, every option and the message together. A small decision head scores each option, and the scores become probabilities. The model is stored as an 8-bit ONNX file and runs with ONNX Runtime Web on your device's CPU.
Licence
Copyright © 2026 Launchables. All rights reserved.
The Myles-22M model, including its weights, configuration and this model card, is proprietary to Launchables and provided only for use through the demo on launchables.ai. It may not be copied, downloaded for reuse, redistributed, modified, reverse engineered or used to train other models without written permission. See the full licence . Organisations can request the weights for internal evaluation on Hugging Face , under the separate Myles Evaluation Licence. For licensing enquiries use our contact form.
Third-party notices
Myles-22M is fine-tuned from cross-encoder/ms-marco-MiniLM-L6-v2 (Microsoft MiniLM), licensed under the Apache License 2.0 ; Launchables modified it by adding a decision head and further training. Training data includes WANLI (Liu, Swayamdipta, Smith and Choi, 2022), licensed under CC BY 4.0, used via the Open-Jev v1.1 dataset. The model runs with ONNX Runtime Web, © Microsoft Corporation, MIT License . These components remain under their own licences.