Most Downloaded AI Models on Hugging Face 2026

What this chart shows. The fifteen model repositories with the most downloads on Hugging Face over the last thirty days, read from the Hub's public API on 21 September 2026 at 08:19 UTC. The sentence embedding model all-MiniLM-L6-v2 leads with 251,048,129 downloads, 2.83 times the 88,600,335 of second-placed ms-marco-MiniLM-L6-v2, and on its own accounts for 33.1 per cent of the 758,585,348 downloads the fifteen drew between them. Eight of the fifteen turn text into vectors for search, ranking or classification rather than writing anything, and only one of the fifteen, Qwen3-0.6B, is a chat-style text generator.

How the data is collected. The Hub counts downloads server-side as it serves files, and sends nothing from the user's machine. Because one repository holds many files, only requests to a set of query files are counted, so a single download is not counted several times: config.json by default, or the file a library nominates for itself, such as pytorch_model.bin or adapter_config.json. Every HTTP request to a query file, GET or HEAD alike, counts as one download. GGUF files are an exception: they are self-contained and tied to no library, so each one is counted, which the documentation warns double counts a whole-repository clone. The diffusers library has its own filter, counting model_index.json together with top-level safetensors, ckpt and bin files. The figure charted here is the downloads field of the API, the same number each model page prints as its downloads for the last month.

Background. The Hub reports an all-time total alongside the monthly one. For the leader it stands at 3,915,587,275 downloads, and for google-bert's bert-base-uncased, fifth here on 46,140,403 in the last thirty days, at 3,238,101,939. Organisations that need more than the public number can use the Hub's Publisher Analytics, which the documentation describes as anonymised request-level logs that let a publisher separate weights from configuration files, exclude CI and CD pipelines, or count unique downloaders. None of that is applied to the figure shown here.

Limits. A download is an HTTP request, not a person: an automated build that pulls a model on every run is counted every time. The window is the last thirty days, so a single large job can move a repository up the list, and the ranking changes from week to week. The query-file rule also differs by library, which makes counts across libraries only roughly comparable. Finally, the list ranks repositories rather than models. The same weights are often published several times, as an original, a quantised copy and a single-file build, and each copy collects its own downloads.