Methodology

Where data comes from

Current model metadata is retrieved from the official Hugging Face Hub API. OpenModelStats does not mirror model cards, files or weights. Everything else on the site — historical observations, interval gains, velocity, rankings — is computed by OpenModelStats from its own recorded history.

Download metrics

• Downloads 30D is a rolling recent-window download counter as reported by the source platform. It is not cumulative.
• Total downloads is the cumulative all-time counter reported by the source.
• Gained 7D / 30D values are calculated by OpenModelStats as the difference in the cumulative all-time counter between two of our own observations taken roughly seven or thirty days apart. We deliberately never derive daily gains from the rolling counter, because that window both adds new days and drops old ones.
• If a baseline observation is missing, or a counter appears to have reset (a negative delta), the affected value is shown as unavailable instead of a wrong number.
• Percentage growth is only displayed when the baseline is large enough for the percentage to be meaningful.

Observation cadence

High-interest models are refreshed more frequently than long-tail ones; every tracked model receives at most one canonical snapshot per calendar day. A page may therefore show a value observed earlier today or within its tier's refresh window. Each chart point represents one day's observation — we do not smooth or interpolate between points.

“Tracked since”

The date OpenModelStats first recorded an observation for a model. Charts begin on that day; no earlier history is back-filled, because it was never observed.

Rankings

Rankings compare tracked models against each other (for example by recent downloads, likes or seven-day gain) and are always labeled by their actual scope, such as “among tracked Text Generation models”. They are not platform-wide ranks unless an exact platform-wide ordering is reliably available from the source.

Momentum (when shown)

Momentum is a composite 0–100 signal used to surface rising models. It blends three percentile-ranked components across eligible tracked models: absolute 7-day download gain (50%), capped 7-day percentage growth (25%), and 7-day likes gained (25%). Percentile normalization prevents tiny models with small absolute numbers from dominating through percentage growth alone. Momentum is only published after at least fourteen consecutive tracked days exist; it is a discovery aid, not a quality score.

Why numbers may differ from the source

OpenModelStats shows the most recently observed value, which can trail the live source by up to a refresh cycle. Source platforms also revise their own counters over time. Small differences between this site and the live platform are expected; when they occur, the methodology above explains which number came from where.

Changes over time

Metric definitions and calculation details may evolve as the product matures. When they do, changes will be reflected on this page so historical context remains interpretable.