Myles vs general-purpose LLMs
A decision deserves its own model.
General-purpose LLMs are built to write. Most decisions inside a product need something else: a probability for every option, fast, cheap and close to the data. Here is how the two compare.
The same 1,400 decisions
| System | Correct | Cost per million decisions | Median time |
|---|---|---|---|
| Myles-400M, on one CPU thread | 91.8% | £4 | 500 ms |
| Hosted LLM, small tier | 86.9% | £900 | 770 ms |
| Hosted LLM, medium tier | 89.6% | £1,880 | 1,390 ms |
| Hosted LLM, large tier | 96.2% | £4,110 | 1,870 ms |
Our own measurement, September 2026, on a frozen set: the hosted models answer zero-shot with one fixed prompt. The large tier is more accurate, at about a thousand times the cost; Myles-G covers the hardest decisions on one GPU. Method and all results
Side by side
| Myles decision models | A general-purpose LLM | |
|---|---|---|
| What comes back | A probability for every option, in one forward pass | Generated text, which your code has to parse |
| Confidence | Calibrated, so a threshold set on held-out data holds; you choose the error rate | A stated confidence, if any, that is not checked against outcomes |
| What it decides | Only what clears its threshold; the rest goes to a person | Everything it is asked, with the same tone either way |
| Where it runs | Your CPU, laptop, phone or one GPU, including offline | Usually the vendor's cloud |
| Where the data goes | Nowhere: it stays on hardware you operate | To the provider, for every call |
| Cost | A licence per deployment, or prepaid credit on our API | Per token, in and out, for every call |
| Lineage | Published: base model, data and licences | Rarely disclosed in full |