Methods
Interconnects' monthly recap of notable open model and dataset releases. This curation is used as the primary source of noteworthy models.
Interconnects' public list of core LLMs on Hugging Face. These models define the checkpoints used for rankings and aggregates. Its manually reviewed parameter and organization-region metadata provide authoritative inputs for Hub sorting, RAM size classes, and geography filters, and it powers other projects such as The ATOM Project.
Historical downloads and RAM analysis are powered by the Interconnects API. Retained download history begins July 5, 2025; coverage details appear below.
Daily token totals for OpenRouter's top-50 most-served models, open and closed, published in their API and synced since January 2025. Provider variants for the same catalog checkpoint are combined; separately cataloged releases remain distinct. Days below the top-50 cutoff are unknown, not zero.
Intelligence Index scores for open and closed models, synced daily from the Artificial Analysis API. They run every evaluation themselves under identical conditions.
Trailing-30-day and all-time download counts and likes from the Hugging Face Hub for every cataloged checkpoint.
Research engineer at Prime Intellect and lead on Artifacts Log.
Founder of a stealth non-profit AI lab; founder and editor of Interconnects AI.
The full architectural parameter count reported for the checkpoint. It is not model file size, a family label, or the number of parameters used for each token.
Tracked-models totals are manually reviewed overrides. Otherwise, the Hub uses a single unambiguous total from artifacts-models sizes; parenthetical active counts are ignored and ambiguous values stay unknown. Supporting links belong in the issue or pull request rather than the registry. Hugging Face API metadata, safetensors totals, and parameter-like model names are never parsed for sorting. The Data control filters the resulting total-parameter ranges; models without a resolved total remain outside selected size ranges.
RAM contextualizes open model downloads across different size categories to identify breakouts, as small models naturally attract more downloads. Therefore, we create different size buckets to calculate the relative popularity for each model in its size group.
RAM compares against the true top-10 downloads for the size class: a score above 1.0× means the model is on track to be a top-10 download ever in its size category.
For release alignment, an exact, manually reviewed release-date correction is authoritative. Without an exact reviewed correction, RAM uses the first usable Hugging Face cumulative-download observation in retained history as the model's analysis start; a Hugging Face repository creation time is not treated as a public launch.
Retained download history begins on . Models with reviewed corrected release dates before that cutoff remain N/A, not zero; RAM does not move their analysis start to a later observation. New models display Pending until seven full days of eligible history are available, then receive an @7d score that is replaced by @30d once that full window is available. Models without required parameter, download, or API metadata display Not scored. Neither Pending nor Not scored is a zero score. Because a full post-release window is required, 30-day coverage effectively begins in early August 2025.
Size-class baselines are recalibrated quarterly on the first day of February, May, August, and November. The reference snapshot is the fixed data cutoff for that quarter, not the date the baseline file was regenerated or deployed. Each stored score records its reference quarter and reference snapshot date so the calculation is reproducible. A RAM refresh can be pending while a model moves from its 7-day to 30-day score or while a quarterly baseline rolls forward; until recalculation succeeds, the last valid score stays visible and is marked RAM refresh pending.
Project VAILA single base model can spawn fine-tunes, merges, and quantizations. Once redistributed, naming conventions do not prove their origin. Behavioral fingerprinting exposes shared lineage—weights, base models, training pipelines—even when those links are undisclosed.
The fingerprinting method runs a forward pass on fixed input tokens, extracts a linear approximation of the model's mapping from the resulting logits, and creates a vector using the magnitude of the corresponding eigenvalues. Fingerprint vectors from two models are compared to produce similarity score between 0 and 1.
The AA column shows each model's Intelligence Index from Artificial Analysis, who run every evaluation themselves under identical conditions. The highest reasoning effort is selected for each model.
Months Behind Frontier calculates the difference between an open model at release and the closed frontier reaching the same score. A value of 3 months means the best closed model first reached the open model's Intelligence Index score 3 months before the open model shipped.
OpenRouter publishes daily token totals for its top-50 most-served models, open and closed, in its API; the Hub has synced them since the dataset's start in January 2025. The OR tok/d column averages a model's tokens per day over the days it charted in the trailing week; model pages also show the peak single-day total and best-ever rank.
Only the daily top-50 is published, so days below the cutoff are unknown rather than zero, and most open models never chart. Provider variants of the same catalog checkpoint are combined; separately cataloged releases remain distinct.