Caveats

• Pulls are Docker Hub’s own counter for the repository, every tag together, since it was created. They count downloads, not users: a CI job that pulls a model on every run counts every time, and older repos have had longer to add up.
• The score is for the model, not the file. Artificial Analysis runs the model at full precision through an API. The 4-bit GGUF that docker model pull fetches by default is a compression of it and will score somewhat lower. The Frontier app’s memory map covers what quantization costs.
• Many scores are estimates. Artificial Analysis has not run its full suite on most small and older models and estimates their Intelligence Index instead; those are marked est. and drawn hollow. Its estimates bottom out near 4.8, so tiny models can look as good as much larger ones.
• Which score belongs to which tag was matched by hand, from each repo’s README and tag names, never from the repo name alone: ai/magistral-small-3.2 holds Mistral Small 3.2, and ai/nemotron3 holds Nemotron 3 Nano Omni. Where a model has a reasoning and a non-reasoning score, the one shown is the variant its tag names, else the reasoning one.
• Embedding models are scored on MTEB, not the Intelligence Index, which measures chat models. The figure is the MTEB leaderboard’s average over the 41 tasks of MTEB English v2 (retrieval, clustering, classification, similarity and more); the multilingual average, over 131 tasks, is shown where the model has run all of them. The two scales have nothing to do with each other, so they sit in separate columns. Reranker, image, video and decision models have neither score.
• Sizes are what Docker Hub reports for each tag (the compressed artifact), and a handful of its tags carry sizes that cannot be the whole model. The size shown is the tag docker model pull fetches by default.