Hugging Face report finds small models dominate downloads as Chinese labs ship the largest open weights
The gap between the models that attract attention and the models that carry real workloads is the lesson for anyone choosing one.
Hugging Face published its State of Open Models: Summer 2026 Observations report on 14 August. Compiled by Adina Yakefu, Apolinario and Irene Solaiman with input from more than 190 contributors, it draws on activity across the Hub between January and August 2026 to describe how open-weight models are being released, licensed and used.
The report says Chinese labs released the largest open models in almost every month of 2026, with parameter counts running from 754 billion to 2.78 trillion, while most US releases stayed under 130 billion parameters; the exception was NVIDIA's 561-billion-parameter Nemotron 3 Ultra. Companies with little prior profile in the field, including Xiaomi, Ant Group and Meituan, crossed the trillion-parameter mark. Of 178 Chinese releases above 20 billion parameters, 59 percent used the Apache 2.0 licence and 22 percent used MIT.
Usage tells a different story. Models under one billion parameters account for 83 percent of all-time downloads, while models above 100 billion parameters make up about 1 percent. The embedding model all-MiniLM-L6-v2 alone has passed 1.55 billion downloads, and only one repository appeared in both the top-25 by downloads and the top-25 by likes. Qwen is the most-built-upon family, with 151,448 derivatives, about 2.6 times Meta's footprint, and Qwen GGUF builds drew 39.6 million monthly downloads against 20.8 million for Gemma and 7.5 million for Llama.
The report also records coding agents as a new class of Hub user. Claude Code accounted for 44.4 percent of identified agent calls in July, down from 67.8 percent in April, while Codex rose from 10.4 to 20.8 percent, and nearly a quarter of July's agent traffic came from unidentified harnesses. Over the same period, public model repositories grew from 2.43 million to 2.96 million and datasets passed one million.
Why it matters
The figures separate two markets that are often discussed as one: the frontier, where scale and headlines live, and the long tail of small models that carry most real workloads. Permissive licences on very large Chinese releases suggest value is shifting toward hosting, hardware and ecosystem position rather than the weights themselves. Agent traffic hitting model repositories directly is an early sign that software, not only people, will be selecting and downloading models. Anyone planning an AI project should expect the practical default to be a small, well-supported model rather than the biggest one available.
The figures separate two markets that are often discussed as one: the frontier, where scale and headlines live, and the long tail of small models that carry most real workloads.
- Models under 1B parameters83%
- Models above 100B parameters1%
Figures: Hugging Face, State of Open Models: Summer 2026 Observations
What you can learn from this
- Parameter count measures capacity, not usefulness. A parameter is a learned weight inside the network, and the total count sets an upper bound on how much the model can represent. That bound matters most for open-ended reasoning and broad knowledge, which is why frontier labs keep scaling. Most production tasks, such as embedding text for search, classifying tickets or extracting fields, are narrow enough that a model with a few hundred million parameters reaches the required accuracy at a fraction of the cost. The download figures reflect the gap between what is impressive and what is sufficient.
- Quantization is what lets big models leave the data centre. Quantization stores weights in fewer bits, for example 4-bit integers instead of 16-bit floats, which cuts memory needs by roughly four times with a modest loss in quality. The GGUF format used by llama.cpp packages a quantized model in a single file that runs on CPUs and consumer GPUs. This is why a trillion-parameter model can be tried on ordinary hardware and why GGUF repositories grew so quickly. The trade-off is that aggressive quantization hurts precision on tasks with tight numeric or factual tolerances.
- Licences decide what you can build, not only what you can read. Apache 2.0 and MIT are permissive licences that allow commercial use, modification and redistribution with attribution. Custom "open" licences often add acceptable-use restrictions, user caps or clauses that revoke rights for some parties. When a family is released under a permissive licence in many sizes, derivative work compounds because each fine-tune inherits the same freedom, which is part of why Qwen's derivative count is so high.
- Downloads and likes measure different things. A download counter mostly records automated pulls from build pipelines, CI jobs and deployed services that fetch a model each time they start. Likes record a human clicking a button, usually in response to an announcement. Only one repository sat on both top-25 lists, which shows that community enthusiasm and infrastructure dependency rarely coincide. Read download counts as a proxy for how many systems would break if a model disappeared.
- Agents are becoming API clients with their own identity. Coding agents send a user-agent string with each request, which is how Hugging Face attributed traffic to Claude Code, Codex and others. Requests without a recognisable identifier are grouped as unknown harnesses. As agents take over tasks such as pulling weights or evaluating checkpoints, rate limits, access tokens and audit logs need to treat them as first-class principals rather than as an extension of whoever launched them.
We teach this
How to use this in practice
- Run a small model locally and measure it. Install llama.cpp or Ollama, pull a GGUF build of a model under four billion parameters such as a small Qwen or Gemma variant, and time a fixed prompt at two quantization levels, for example Q4 and Q8. Record tokens per second, memory use and whether the answers differ on ten of your own test questions. Done looks like a one-page table comparing size, speed and quality, so you can justify a model choice with your own numbers rather than a leaderboard.
- Audit the licences of every model your code loads. Search your repositories for
from_pretrained(,hf_hub_downloadand hard-coded model IDs, then open each model card and note the licence field. Flag anything that is not Apache 2.0, MIT or a licence that has already been reviewed for your project. Done means a short inventory listing model ID, licence and where it is used, with each non-permissive entry assigned to someone to resolve. - Pin model revisions instead of pulling the default branch. Hugging Face repositories are git-backed, so every file has a commit hash. Change each download call to pass a
revisionargument with the specific hash, and store that hash next to your code. Done looks like a fresh clone of your project producing byte-identical model weights on a second machine, verified by comparing file checksums. - Give any script or agent that hits a model hub its own identity. If you have automation that downloads models, configure the HTTP client to send a descriptive user-agent such as the project name and version. Check the hub's terms for rate limits and create a scoped access token for the automation rather than reusing a personal one. Done means the hub's access logs, or your own proxy logs, show the automation under its own name, and revoking its token would not disturb anyone else's access.
Sources
- State of Open Models: Summer 2026 Observations — Hugging Face
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report