Myles decision models

One job, done well, on any machine.

Myles reads a situation, a question and the possible answers, and returns a probability for every option in a single pass. It classifies, routes, checks and approves: the decisions inside every AI product.

Myles sitting on a rock, checking his phone, beside a server rack

91.8%correct on 1,400 held-out decisions, ahead of the medium tier of a leading hosted LLM family

1/450thof the cost per decision, on a single CPU thread

0GPUs, cloud calls or data leaving your network

Myles-400M on our frozen evaluation set. Method below.

The family

Model	Runs on	Parameters	Typical speed
Myles-22M	Your browser or phone	23M	27 ms
Myles-150M	A laptop	149M	196 ms
Myles-400M	The cheapest server you have	395M	about 500 ms
Myles-L	One 24 GB GPU	14.7B	97 ms
Myles-1B	In training	1B	—

Median time for one decision on one CPU thread of a desktop Intel processor: 22M and 150M as 8-bit builds, both timed while the machine was busy with other work (an idle machine is faster), 400M at full precision (its int8 build is about three times faster). Myles-L on one desktop GPU in 4-bit. It is the escalation tier: an adapter on an open 14B base that reads up to 4,096 tokens and takes the cases the CPU models are unsure of. That base was pre-trained partly on synthetic text from hosted models. Each model's card records its full lineage.

How it compares

Every model answers the same 1,400 decisions, frozen before any comparison: 800 from an open decision dataset (half of them from a shifted distribution), 300 product tasks and 300 everyday tasks. The hosted models answer zero-shot with one fixed prompt.

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

Myles-400M leads the medium tier by 2.2 points (95% confidence interval 0.4 to 4.0 points) at about one 450th of its cost. Myles is trained for these decisions; the hosted models answer with one fixed prompt. Local cost assumes a cloud CPU at about 3p per hour; hosted prices are list prices converted at £1 = $1.33; hosted times include the network.

Reasoning it was never shown

74%correct on multi-rule problems from domains it never trained on, where chance is 16%

59.7%on an independent public benchmark for decision models, the highest of the open models under 1B parameters we have tested

77.9%on the same benchmark for Myles-L in one forward pass on one GPU

Latest Myles-400M build, September 2026, and Myles-L, October 2026. The reasoning problems are generated with exact answers, and whole domains are held out of training. The benchmark score covers its 231 public tasks, scored with the benchmark's own code.

Confidence you can certify

Myles keeps a decision when its confidence clears a threshold set on held-out data, and hands the rest to a person or a larger model. Each kind of input gets its own threshold. At a 5% target error, the Myles-400M build behind the 91.8% result keeps 73% of decisions, with 2.7% error on unseen test data. At 2%, it keeps 40% with 1.4% error, and at 1%, 16% with 0.5%. Every result lands inside its target.

Organisations can request the Myles-400M and Myles-L weights for internal evaluation on Hugging Face, under the Myles Evaluation Licence.

Early access, evaluation licences and pilots.