Local model reach is not the same as Hub use

Local reach and measured use answer different operating questions. Community software can spread a trillion-parameter model across consumer computers. Hugging Face still sees almost all Hub download volume below that scale. A routing plan needs both facts without turning either into a claim about total market adoption.

Hugging Face published its report on August 14, 2026. Its period analysis covers the first seven months of 2026. Hugging Face is not a Muniment customer or endorser.

Small models carry the measured volume

A parameter is one learned weight inside a model. Its count labels model size. It does not score quality.

Hugging Face reports the following split among models that declare a parameter count.

These are Hub counts with declared-parameter filters, not a census of model work. Small, stable models can sit inside scheduled pipelines and accumulate repeated downloads. A large model can arrive through another channel or run behind an interface that creates no Hub download for each request.

Reach sets the option, use records the path

Hugging Face says llama.cpp can spread a trillion-parameter model across a few consumer machines. That result expands the set of models a local route can reach. It does not show how often operators choose that route.

The distinction matters in production. Feasibility can justify a local tier, while measured demand decides whether that tier deserves capacity and support. A download share cannot settle either decision by itself.

Hugging Face states that downloads measure activity inside the Hub ecosystem. They exclude API use, private deployments, and models distributed through other channels. The figures therefore describe Hub activity, not overall model use or market share.

Treat reach as an available route and use as an observed event. If you need records that keep those facts separate, join the waitlist.