On August 14, 2026, Hugging Face published a new analysis of open-model activity. Only one repository appeared in both its top 25 model repositories by downloads recorded during the report’s 2026 measurement window and its top 25 by likes.
The mismatch is a useful warning: a repository that attracts attention and one that generates heavy retrieval traffic are not necessarily the same thing.
Downloads and likes reflect different behavior
None of the models published in 2026 reached Hugging Face’s download top 25, while 13 of the 25 most-downloaded repositories dated to 2022.
Hugging Face also reported 1.55 billion downloads over seven months for sentence-transformers/all-MiniLM-L6-v2, compared with 5,156 likes. The company interprets that contrast as evidence that likes tend to reflect attention around releases, while downloads accumulate around models repeatedly retrieved by software workflows.
That interpretation is plausible, but the figures do not reveal how many distinct people or production systems used the repository.
The broader distribution is also highly concentrated. Hugging Face reports that 85.6% of model repositories have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.
What the download counter measures
Hugging Face’s download-count documentation says the Hub counts requests for designated query files in a model repository. Every qualifying GET or HEAD request contributes to the counter, and the calculation happens on the server as files are served.
The query file varies by library. A request for config.json may count under the default method, while some integrations use files such as pytorch_model.bin or adapter_config.json.
This makes the metric useful for measuring retrieval activity, but it is not a unique-user count. Repeated requests from automated systems or CI/CD jobs can contribute multiple times. Hugging Face directs publishers to more granular request logs when they need to exclude CI/CD traffic or deduplicate downloaders.
A public download total therefore does not establish how many people used a model, how many distinct applications depend on it, or how much production workload it serves.
How to read model-adoption claims
Likes are best treated as an attention signal. Download counts measure qualifying retrieval requests. Repeated requests may be consistent with a model being embedded in software dependencies, but the public counter cannot prove that explanation on its own.
Production adoption requires different evidence, such as deduplicated telemetry, deployment records, customer disclosures or measured workload volume.
No single metric answers every question. A careful comparison should state the measurement window, explain the counting method and avoid translating pulls into users or deployments. That distinction matters as model developers and infrastructure providers compete to describe whose ecosystem is gaining traction.
