Decision models
The decision layer for the enterprise
Small, calibrated decision models that run on the CPUs you already own, inside your network, with an error rate you sign off before go-live.
Make the routine decisions inside your services faster, cheaper and safe to sign off.
Every option gets a probability; there is no text to generate and nothing to parse.
The threshold is set on held-out data before anything goes live.
Anything below its threshold goes to a person, with the probabilities attached.
On your servers, laptops or one GPU. Nothing is sent to us.
State
Charged twice for order 4471, need the extra payment back
Question
Which team should take it?
- Billing0.93
- Returns0.04
- Other0.02
- Technical0.01
Myles. One family, one job.
| Model | Runs on | Median decision | Status |
|---|---|---|---|
| Myles-G 31B | One data-centre GPU | 48 ms | Available |
| Myles-G 12B | One workstation GPU | 31 ms | Available |
| Myles-G Edge | Laptop- and phone-class GPUs | 21 ms | By request |
| Myles-400M | The cheapest server you have | About 500 ms | Available |
| Myles-150M | A laptop | 196 ms | Available |
| Myles-22M | A browser or a phone | 27 ms | Available |
Median time for one decision: Myles-G on one GPU, the others on one CPU thread of a desktop processor.
Why enterprises choose Mellix
- Network policy
- Inboundfrom your servicesAllowed
- Outboundto the internetBlocked
- Outbound to Mellix0 bytes
Your data stays put.
Our models run on hardware you operate, including air-gapped networks.
- Go-live sign-off
- Error target5%
- Keptheld-out data73%
- Wrongof those kept2.7%
Error rates you sign off.
A certified error rate on held-out data, set before go-live.
- Lineage
- TrainedUK-owned hardware
- EvaluatedFrozen held-out set
- MeasuredSeptember 2026
Built in the UK.
Trained and evaluated on UK-owned hardware, with published lineage.
- Pricing
- Self-hostedper deploymentOne licence
- Hosted APIper million decisions~£1.05
- Output tokensFree
A fixed cost.
A licence per deployment, or prepaid credit on the Decision API.
- 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