launchables°
01 · Decide
One pass, every option scored.
Myles reads the situation, the question and the options, and gives each option a probability in one forward pass. No text to generate, nothing to parse.
Certified decision models · Built in the UK
Decisions you can sign off.
Myles decision models give every option a probability in a single pass, keep only the decisions that clear a certified error rate, and pass the rest to a person. Call them through our API, or run them on hardware you own.
- situation
- Customer email: "I was charged twice for my March invoice. Please refund one."
- question
- Which team should handle this?
- options
- billing, technical support, sales, complaints
Kept: above the certified threshold
91.8%correct on 1,400 held-out decisions, ahead of the medium tier of a leading hosted LLM family
£4per million decisions on one CPU thread, against £1,880 for that hosted tier
74%on multi-rule problems from domains it never trained on, where chance is 16%
0cloud round trips: every decision is made inside your network
Myles-400M on our frozen evaluation set, September 2026. Every model answers the same 1,400 decisions; the hosted models answer zero-shot with one fixed prompt. Method and full results
How it decides
Every decision is kept or handed over.
A probability is only useful if you can trust it. Myles is calibrated, so a threshold set on held-out data holds on new decisions, and you choose the error rate before anything goes live.
02 · Certify
A threshold you can sign off.
Each kind of input gets its own threshold, set on held-out data, so the decisions Myles keeps stay inside the error rate you chose.
03 · Hand over
The rest goes to a person.
Anything below its threshold goes to a person or a larger model, with the probabilities attached. Nothing is forced through.
Kept by MylesHanded to a person
Myles-400M on unseen test data, September 2026: at each target error rate, the share of decisions it keeps and the error on those it keeps. Every result lands inside its target. All results and the method
Text and images
Decisions that turn on a picture.
Myles-G 12B now decides from images as well as text. Send the receipt, the screenshot or the product photo with your question, and get the same calibrated yes or no, choice or score back, with a probability for every option, in one API call.
Approve this receipt?
Check a receipt photo against the expense claim and your policy before it is paid.
Is it safe for the agent to tap this?
A guard for computer-use agents: check the screen before the agent acts, and hand the risky taps to a person.
Fit to ship?
Visual inspection from a product photo: pass it, or send it for a second look.
Which item is the target?
Pick the right object in an image from your list of options, with a probability for each.
Internal measurement on 934 held-out images across receipts, app screens, product photos and object boxes, 16 Oct 2026; not an official benchmark placement. Speed measured the same day. The figures are for Myles-G 12B; Myles-G 31B reads images too.
Deployment
Runs where your data already lives.
On your servers
CPU only, including disconnected networks. Myles-400M makes a decision in about half a second on one thread.
On laptops and phones
Small enough for the devices your teams already carry. Myles-22M decides in 27 ms, even in a browser.
On one GPU you own
The most accurate models, on a single workstation or data-centre GPU. Available by request.
Where it is used
The decision points inside your services.
Is this message a scam? Which team should take it? Is this answer supported by its source? Can this agent issue that refund? Small, frequent, consequential decisions are where Myles sits.
Public sector
Triage, eligibility checks and casework routing, on your own estate, with a lineage your risk owner can inspect.
Learn more
Financial services
Scam and fraud signals, complaint classification and routing, with an error rate set before it goes live.
Learn more
Health and care
Referral triage, personal-data checks and message routing, without patient data leaving the building.
Learn more
AI product teams
Guardrails, routing and grounding checks inside your agents, in milliseconds, at a fixed cost.
Learn moreThe Myles models
One family. One job: the decision.
| Model | Median decision |
|---|---|
| Myles-G 31BOne data-centre GPU | 48 ms |
| Myles-G 12BOne workstation GPU | 31 ms |
| Myles-G EdgeLaptop- and phone-class GPUs | 21 ms |
| Myles-400MThe cheapest server you have | About 500 ms |
| Myles-150MA laptop | 196 ms |
| Myles-22MA browser or a phone | 27 ms |
Median time for one decision: Myles-G on one GPU, the others on one CPU thread of a desktop processor.
Latest
What we are shipping.
- Myles-G 31B on the Decision APIOur most accurate model, for every account: £0.005 per million input tokens, output free, text and images.Release
- Myles-G 12B decides from imagesReceipts, screenshots and product photos, in the same request as the text, with a probability for every option.Release
- Myles-G 12B on the Decision API: $0.002 per million input tokensOur first published API price, with output tokens free.Release
- Certified error rates, explainedHow a calibrated model decides what it keeps, and why the threshold is set before go-live, not after.Note
- The Decision API, version 1Three question types, a probability for every option, several questions about one situation in one request.Release
- Myles-400M: 91.8% on 1,400 held-out decisionsAhead of the medium tier of a leading hosted LLM family, on one CPU thread, at about one 450th of the cost.Evidence
Licensing
Evaluate. License. Run it yourself.
01 · Evaluate
Start with your own decisions.
Send us a sample. We measure Myles on it and show you the error rate before you commit.
02 · License
One licence per deployment.
Monthly or annual. No per-decision fees: it runs on your hardware.
03 · Run
Upgrade with the evidence.
Every new version ships with its measured results, so you can compare before you switch.
How licensing works Decision API: from £0.0015 per million input tokens
Start here
See what Myles does with your decisions.
Get a key and call the Decision API, or run Myles on hardware you own.