Caveats

• A download is not a user. Hugging Face counts a download each time a model’s config file is fetched, over the last 30 days. Test suites, CI jobs and libraries that load a default model on import fetch the same small models again and again, which is why years-old models like all-MiniLM-L6-v2 and gpt2 sit near the top. The all-time count is shown beside it.
• Copies are counted once, under the original. A repo Hugging Face tags as a quantization of another (base_model:quantized:…) is folded into that model’s row, whoever uploaded it, and its downloads are added to the total. Each original’s 300 most-downloaded copies are counted; the long tail beyond adds little. Fine-tunes, merges and adapters change the weights, so they keep rows of their own and borrow no score.
• Two ways onto the page. Most rows are the most-downloaded models with a given Hugging Face task label: text-generation and image-text-to-text (chat), sentence-similarity and feature-extraction (embedding), text-ranking (reranker), automatic-speech-recognition, text-to-speech and text-to-image, and, for decision models, which have no task label of their own, the decision-model tag, keeping the top by total downloads. The rest are on the page for their score, however few download them: every chat model Artificial Analysis has measured (not estimated), the best embedders on MTEB English, every open model on the two speech leaderboards, and every open model on the Decision Index board. Leaderboard entries with no public weights (, such as API-only embedding and speech services) are left out. Feature-extraction models count as embedders only if they use sentence-transformers or MTEB has scored them, which keeps out audio and image encoders. Test fixtures ( repos from accounts named for testing, or with no weights) are left out.
• Each kind has its own score, matched by exact repository, never by name. Chat models: the Artificial Analysis Intelligence Index, which records the Hugging Face repo holding each open model’s weights. Embedding models: MTEB English v2, which files its results under repo ids. Speech to text: the Open ASR Leaderboard’s average word error rate on English, lower is better. Text to speech: the Artificial Analysis Speech Arena’s Elo, from blind listening votes, which links each open model to its repo. Decision models: the community Decision Index, which runs every entrant on the same GPU, links each to the repo it ran, and runs the closed TypeSafe Jev alongside; the headline blends its public benchmarks with private ones, and makers’ own numbers (Liquid’s d1 launch figures, for one) are not used. The scales have nothing to do with each other, so the Score column sorts by benchmark first and the map shows one at a time. A model a benchmark has not run shows no score. Many Artificial Analysis scores for small and older models are estimates, marked est. and drawn hollow; where it scores the same weights in more than one mode (reasoning and not), the headline is the best measured mode.
• The score is for the original, not its copies. A 4-bit GGUF of a model will score somewhat lower than the full-precision weights Artificial Analysis ran.
• Sizes come from each repo’s file list, not its storage total. Hugging Face reports how much a repo stores, which counts every format it ships: all-MiniLM-L6-v2 stores 1.9 GB for 91 MB of weights. Weights is the original’s top-level Safetensors files (else its PyTorch ones), at whatever precision it was published in. 4-bit GGUF is the Q4_K_M build (else the nearest 4-bit one) in the most-downloaded GGUF repo for the model, split parts summed and vision projectors left out; a model with no GGUF copy shows none.
• Rerankers and image generators have no score here. There is no common benchmark rerankers are all run on, and Artificial Analysis’s image arena loads its table in the browser rather than publishing it in the page, so there is nothing this page can read without guessing.