TL;DR
A Hugging Face report covering January through August 2026 finds that Chinese laboratories increasingly lead releases of very large open-weight models, while US activity is shifting toward hardware and infrastructure companies. The data also shows a sharp divide between models attracting attention and the smaller, older systems developers actually use.
Chinese AI laboratories increasingly set the size ceiling for frontier open-weight models in 2026, while US open-model activity has shifted toward hardware and infrastructure companies, according to a Hugging Face analysis of activity from January through August. The report also found that new frontier releases attract attention but little broad adoption, with downloads still dominated by older, smaller models embedded in production systems.
Hugging Face said the largest Chinese open model released in nearly every month of the period was bigger than any original model published by a US laboratory. The Chinese monthly ceiling ranged from 754 billion to 2.78 trillion parameters. The US ceiling remained below 130 billion in five of seven measured months, apart from releases including Thinking Machines Lab’s 952-billion-parameter Inkling and NVIDIA’s 561-billion-parameter Nemotron 3 Ultra.
The report identified two Chinese publishing strategies. Moonshot, MiniMax, Xiaomi and Z.ai focused largely on models above 70 billion parameters, while Tencent and Alibaba’s Qwen released models across a much wider size range. Hugging Face said community-produced quantizations often make very large releases runnable on less powerful hardware within days, reducing the need for laboratories to publish smaller versions themselves.
US participation remains substantial but has changed form. AMD and NVIDIA each published more than 200 model repositories during the period, ahead of other organizations, while Liquid AI released about 100. Hugging Face characterized much of this work as conversion, optimization and hardware support, rather than the creation of original frontier-scale foundation models.
State Of Open Models: Summer 2026 Observations
Chinese laboratories are setting the size ceiling for frontier open-weight releases. US activity is shifting toward hardware and infrastructure, while the models developers actually depend on remain smaller, older and deeply embedded in production.
China raises the size ceiling
In nearly every measured month, the largest Chinese open model exceeded every original model published by a US laboratory. The split is not simply about volume: it reflects different publishing strategies and a changing division of labor.
Scale-first portfolios
Moonshot, MiniMax, Xiaomi and Z.ai concentrated largely on models above 70 billion parameters.
>70BOne family, many sizes
Tencent and Alibaba’s Qwen released across a wider size range, giving developers more deployment choices.
Wide spectrumQuantization closes the gap
Community conversions can make huge models runnable on less powerful hardware within days of release.
Days, not monthsObserved parameter scale
Relative to 2.78T maximum*The US ceiling stayed below 130B in five of seven measured months. Parameter count indicates scale—not quality, efficiency, safety, reliability or demand.
The models people notice are not the models systems use
Likes capture excitement around new releases. Downloads accumulate as models are retrieved by applications, tests, builds and automated pipelines. The two signals answer different questions.
| Signal | What it reveals | Frontier releases | Older small models | Key limitation |
|---|---|---|---|---|
| Likes | What the field is excited about now | ✓ | ~ | Short-term attention can fade quickly |
| Downloads | What workflows repeatedly retrieve | ✗ | ✓ | Not a direct count of users or deployments |
| Parameters | The nominal scale of a model | ✓ | ✗ | Does not establish performance or efficiency |
| Repository count | Breadth of published artifacts | ~ | ~ | Conversions and optimizations inflate totals |
1.55 billion downloads
The older all-MiniLM-L6-v2 model reached this total in seven months while recording 5,156 likes—a vivid example of quiet, production-level dependence.
Likes
Downloads
Just one repository appeared on both lists. Thirteen download leaders dated to 2022; none published in 2026 entered the download top 25.
“Likes are the right instrument for reading what the field is excited about, downloads for reading what it currently depends on.”
Hugging Face report
Growth masks extreme concentration
The public inventory expanded rapidly from January through August, yet most repositories remained barely downloaded. Publishing activity and actual usage are moving on very different curves.
New public model repositories across the measured period.
Dataset growth broadened the Hub’s supporting ecosystem.
More demos and applications made models easier to explore.
of models had fewer than 200 lifetime downloads.
A very long low-usage tailof repositories generated 99.2% of all downloads.
A tiny group carries almost all usageDownloads are a signal, not a census
A retrieval may come from production, testing, an automated build or repeated caching behavior. Download totals indicate dependence, but they do not reveal the exact number of users, deployed applications or business outcomes.
From frontier release to production reality
The next phase will test whether 2026’s headline models become durable development foundations—or remain admired artifacts while established smaller systems continue doing the work.
🧪 Frontier release
A laboratory publishes a large open-weight model and establishes a new scale marker.
⚙️ Optimization
Hardware firms and community developers convert, quantize and tune the release.
🧩 Integration
Developers test cost, speed, reliability and compatibility inside real workflows.
📈 Sustained use
Repeated downloads reveal whether the model becomes durable infrastructure.
Will frontier downloads persist?
Watch whether the attention surrounding 2026 releases converts into sustained retrieval and integration.
Will Qwen become a common base?
Its broad model range may make the family useful across more hardware profiles and development contexts.
Will US labs return above 100B?
Later releases could change the regional size ranking established during this partial-year window.
Will infrastructure remain the US engine?
AMD and NVIDIA each published more than 200 repositories, much of it focused on conversion, optimization and hardware support.
What is the main finding?
Chinese laboratories led most frontier-scale open-model releases, while US publishing activity leaned more heavily toward hardware and infrastructure companies.
Are the newest models the most used?
No. None of the models published in 2026 entered the top 25 by downloads; older, smaller models still dominate established pipelines.
Why do likes and downloads diverge?
Likes capture attention around a release. Downloads accumulate through repeated retrieval by applications, tests and automated systems.
Does larger mean better?
No. Parameter count describes scale, not performance, cost, adoption, safety or reliability. Each requires separate evidence.
Adoption Favors Stable Small Models
The findings challenge the use of popularity signals as evidence of real-world adoption. Among the 25 most-downloaded repositories and the 25 repositories with the most likes, only one appeared on both lists. No model published in 2026 reached the download top 25, while 13 entries dated to 2022.
That gap matters because likes often capture short-term interest in new frontier releases, while downloads indicate models running repeatedly inside applications and automated pipelines. The older all-MiniLM-L6-v2 model recorded 1.55 billion downloads in seven months despite receiving 5,156 likes, according to the report.
As an affiliate, we earn on qualifying purchases.
Hub Growth Masks Heavy Concentration
The Hugging Face Hub continued expanding during the measured period. Public model repositories increased from 2.43 million to 2.96 million, datasets rose from 711,000 to 1 million, and Spaces grew from 1 million to 1.44 million.
Usage remained highly concentrated beneath that growth. About 85.6% of models had fewer than 200 lifetime downloads, while 1.5% of repositories generated 99.2% of downloads. This means repository growth alone does not show how broadly models are being used.
“Likes are the right instrument for reading what the field is excited about, downloads for reading what it currently depends on.”
— Hugging Face report
large open AI models hardware support
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Downloads Leave Usage Questions Open
The data does not establish how every downloaded model is used. A download can come from a production service, testing workflow, automated build or repeated retrieval, so download totals are an adoption signal rather than a direct count of users or deployed applications.
It is also unclear whether the US-China release gap will persist. The observations cover part of 2026, and later releases could change the size rankings. Parameter count alone also does not establish model quality, efficiency, safety or commercial demand.
As an affiliate, we earn on qualifying purchases.
Release Strategies Face Production Tests
Attention will now turn to whether 2026 frontier models gain sustained downloads, whether Qwen’s broad model range becomes a common development base, and whether US laboratories resume publishing original models above 100 billion parameters. Future Hub data should also show whether hardware-optimized releases remain the main source of US open-model growth.
INFINIBAND FOR HIGH-PERFORMANCE COMPUTING AND AI CLUSTERS: Configure RDMA networking, optimize GPU interconnects, and build low-latency infrastructure for distributed training and HPC workload
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main finding of the Hugging Face report?
Chinese laboratories led most frontier-scale open-model releases during the measured period, while US publishing activity was concentrated more heavily among hardware and infrastructure companies.
Are the newest open models also the most widely used?
No. Hugging Face found that none of the models published in 2026 entered the top 25 by downloads. Older, smaller models continue to dominate usage because many are embedded in established software pipelines.
Why are likes and downloads so different?
Likes generally reflect attention around a release, while downloads accumulate when models are repeatedly retrieved for applications, testing or automated systems. Each metric records a different form of activity.
Does a larger parameter count mean a better model?
No. Parameter count describes model scale, but it does not by itself prove stronger performance, lower operating costs, wider adoption or greater reliability. Those outcomes require separate evaluation and deployment evidence.
Source: Hugging Face
