Methods
How models in the Artifacts Hub are selected, tracked, measured, fingerprinted, and compared.
Interconnects' monthly recap of notable open-weight releases. Models enter the Hub through these editorial batches.
Interconnects' public list of core LLMs on HuggingFace. These models are the checkpoints we track downloads and RAM scores for, and it powers other projects such as The ATOM Project. Additional multimodal models are included in this site via their coverage in the Artifacts Log series.
Historical data and analysis for core models is powered by the Interconnects API.
Research engineer at Prime Intellect and lead on Artifacts Log.
Founder of a stealth non-profit AI lab; founder and editor of Interconnects AI.
RAM contextualizes open model downloads across size categories to identify which models are genuinely breaking through in their class. Small models dominate raw download counts, so comparing a 4B model against a 400B model directly is misleading. RAM normalizes by bucket.
A score of 1.0× means a model is tracking to be a top-10 download in its size category. Scores are computed at milestones: 7, 14, 30, 60, 90, 180, and 365 days post-release. Median is used instead of mean to avoid outlier distortion — a single viral model can skew averages by nearly 10×.
Project VAILA single base model can spawn hundreds of fine-tunes, merges, and quantizations. Once renamed and redistributed, naming conventions alone can't tell you where a model came from. Behavioral fingerprinting answers that question from the weights themselves.
The method runs a forward pass on fixed input tokens, extracts a linear approximation of the model's mapping from the resulting logits, and reduces it to a unit vector via eigenvalue decomposition. Two models are compared with L1 distance, producing a 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 — standardized zero-shot prompts, sandboxed code execution, fixed temperature and token limits — so scores are directly comparable across open and closed models. Agentic tasks (GDPval-AA, τ³-Bench) carry the largest weight, followed by coding (Terminal-Bench, SciCode) and scientific reasoning (HLE, GPQA Diamond).
We cache the score from the Artificial Analysis API once daily and match it to hub models by name. Models AA hasn't benchmarked — most dedicated tools and older releases — show no score. Where AA lists separate reasoning-effort variants, the default configuration's score is shown.
Frontier lag places general-purpose open models on the closed frontier — the best-score-over-time curve of closed models from OpenAI, Anthropic, Google, and xAI. A lag of 3 months means the best closed model first reached the open model's Intelligence Index score 3 months before the open model shipped.