Generating an answer is the long way round
Ask a chat model which team should handle a message and it writes a reply, one token at a time. Then your code has to find the answer in that text, cope when it is phrased differently, and guess how sure the model was. Each step costs time and money, and the confidence is an afterthought.
Scoring the options is the short way
Myles takes the situation, the question and the options together and returns a probability for every option in a single forward pass. There is no text to generate and nothing to parse. You get the full picture: not only that billing is most likely, but that complaints is the runner-up at 4%.
No text to generate, nothing to parse: one pass, every option scored.
Why it matters
- Speed. Myles-22M decides in about 27 ms in a browser; Myles-400M in about half a second on one CPU thread; Myles-G in tens of milliseconds on one GPU.
- Cost. On our frozen comparison, Myles-400M on one CPU thread costs about £4 per million decisions, against £1,880 for the medium tier of a leading hosted LLM family.
- Trust. Because the model scores options directly, its probabilities can be calibrated and certified. That is what lets you automate some decisions and hand over the rest.
- Structure. Three question types cover most decisions: yes or no, one of several labels, or a level on a scale.
When to use something bigger
Decision models are not for writing letters or holding a conversation. They are for the many small decisions inside a system that does. Where a decision is too hard for a small model, Myles hands it to a person or to a larger model, with its probabilities attached.