Hugging Face separates open-model attention from actual adoption
Hugging Face’s summer 2026 report shows that media attention and real adoption of open models barely overlap anymore, and that Chinese labs dominate the frontier by sheer size. Small models and Qwen remain the practical layer, while agents become the Hub’s primary user.
14 August 2026. Hugging Face publishes its biannual “State of Open Models” report. 2.96 million. That is the number of public model repositories on the Hub, up from 2.43 million at the start of the year. One. That is how many repositories appear in both the top 25 by downloads and the top 25 by likes, the bluntest measure of the divorce between attention and adoption.
The report covers the first seven months of 2026 and delivers a picture of the open-model ecosystem that contradicts several assumptions. The most persistent one: that the race is decided purely on frontier model performance. It is decided mostly on distribution, hardware and ecosystem position.
The frontier has moved east
The most striking figure is a monthly ceiling. In almost every month of 2026, the largest open model from a Chinese lab was bigger than any model an American lab released. The Chinese monthly ceiling ranged between 754 billion and 2.78 trillion parameters, while US models stayed under 130 billion in five of seven months, with two exceptions: NVIDIA’s Nemotron 3 Ultra (561 billion) and Thinking Machines’ Inkling (952 billion).
That size is no longer an isolated feat. Xiaomi, Ant Group and Meituan all cleared a trillion parameters this year, and none was a household name in open weights twelve months ago. And a lab no longer has to ship a small model to be reachable: the community quantization layer makes a large one runnable within days.
Two camps emerge. Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70 billion: a developer’s first encounter with them is a model too large to run on their own hardware. Tencent and Alibaba Qwen cover the whole range instead. These are two rational bets on different prizes: benchmark position versus family standardization.
Attention and adoption are two different economies
The report’s most instructive demonstration fits in one line: of the 25 most-downloaded repositories this year and the 25 most-liked, one appears in both lists. Downloads are counted inside the window rather than lifetime, so nothing is credited for merely existing longer.
Two numbers capture the gap. all-MiniLM-L6-v2, a 2022 embeddings model, was downloaded 1.55 billion times in seven months against 5,156 likes. Moonshot’s Kimi-K3 was downloaded about 60 times per like received. No model published in 2026 reaches the download top 25, while thirteen of the twenty-five date from 2022.
The two instruments record different acts. A like says a release matters, and goes to frontier models in the weeks after they ship. A download says something is wired into a pipeline that runs on a schedule, and accrues to small, stable models over years. Confusing the two is the most common mistake in Hub coverage.
Weights are no longer a licensing business
If frontier models were a licensing business, the biggest releases would carry the tightest terms. The opposite holds. Of 178 Chinese releases above 20 billion parameters this year, 59% carry Apache 2.0 and 22% carry MIT, and almost none impose non-commercial restrictions. On the American side of the same band, 29% are Apache or MIT, 41% sit under custom terms and 30% declare nothing at all.
DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT. Whatever these releases are for, it is not license revenue: the return comes from API and cloud business, hardware positioning, or the ecosystem position itself.
The hardware proof is the clearest. The two organizations publishing the most new open models this year are AMD and NVIDIA, each above 200 repositories, far ahead of the rest of the field. A model optimized for your silicon, freely available, is the clearest proof that the silicon works.
Qwen has become the base model
Ecosystem position is measured not by a lab’s own releases but by what the community builds on top. By that measure Qwen dominates: Qwen-based models account for 151,448 derivatives on the Hub, 2.6× Meta’s total footprint and 4.7× the Llama repositories specifically. Google follows with 82,506 derivatives. The flow runs at 180 to 210 new repositories per day across the first seven months.
That position was built largely by the community: the derivatives are downstream work, not Qwen releases. Even among the 28,531 GGUF conversions of Qwen models, the lab published only 54. Consistent releases, size coverage and the Apache 2.0 license reinforce each other: a broad family attracts developers, who build more derivatives, which make the ecosystem more attractive.
Small models remain the practical layer
Downloads tell a different story from the headlines. Among models that declare a parameter count, those under 1 billion take 83% of all-time downloads; everything above 100 billion takes 1%. Restricting to 2026 changes nothing: 3% of volume goes to models above 70 billion.
So how does a trillion-parameter model reach anyone at all? Through llama.cpp. The ceiling moved with it: the July snapshot carries GGUF builds of DeepSeek-V4-Flash at 284 billion parameters and Kimi-K3 at 2.8 trillion. Local inference no longer means an 8-billion-parameter model on a laptop, but a trillion-parameter mixture-of-experts spread across a few consumer machines. That route runs on Qwen: 39.6 million GGUF downloads a month, nearly twice Gemma (20.8 million) and more than five times Llama (7.5 million).
The runtime layer is also growing faster than the model core. Repositories declaring the gguf library rose 464%, lerobot 194%, mlx 148%, against 16% for transformers. The layer that decides where a model can physically run, local inference formats, Apple silicon, robot control stacks, is growing three to seven times faster than the core.
Agents are the new reader
The newest chapter could not have been written in March: the instrument did not exist. The agent-usage dataset, published in July, records the agent/<name> token that coding agents send when they call the Hub. Claude Code led July with 44.4%, but a single month hides the real trend: it held 67.8% in April while Codex climbed from 10.4% to 20.8%. This is a market with no incumbent, where one release or one changed default can move half the traffic in a month.
The report closes on an incident Hugging Face documented itself: in July, what appears to be the first documented case of an autonomous agent running a sustained intrusion on its own initiative happened on its systems. The closed models called in to analyze the attack code declined the work on safety grounds; the analysis was completed on a quantized open model, GLM-5.2, running on house infrastructure.
Verdict
If you are choosing a model to deploy in production, look at downloads, not likes: they measure what is wired into running pipelines, not what excited the community for two weeks. A proven small model often beats a recent frontier release for a stable workload.
If you are standardizing a family to fine-tune or deploy, the Qwen signal is hard to ignore: 151,448 derivatives and a daily conversion flow speak to a base layer that has become infrastructure. And if your roadmap intersects with agents, treat them as a first-class user: they are the fastest-growing reader, and the one for whom the MCP consolidation into the Linux Foundation matters most.