Hugging Face just dropped their summer 2026 state of open models report, and if you're paying attention to where the actual frontier of open-weight AI lives, the numbers tell a story that should make everyone recalibrate.
The TL;DR: China now owns the top end of open-weight model releases. American labs have largely ceded that ground, pivoting instead to infrastructure, hardware optimization, and closed flagship models. And the community has quietly standardized on Qwen as the base model family that actually matters.
Let me walk through what the data shows, because the details matter.
The Frontier Ceiling Is Now Chinese
In almost every month of 2026, the largest performant open model from a Chinese lab exceeded anything American labs released on their own. China's monthly ceiling ran between 754B and 2.78 trillion parameters. America's ceiling stayed under 130B in five of seven months.
The exceptions were NVIDIA's Nemotron 3 Ultra at 561B and Thinking Machines' Inkling at 952B. But here's the kicker: most U.S. releases above 100B parameters weren't original models—they were builds on top of Chinese base models.
This isn't about capability in the abstract. It's about what labs are choosing to release into the open ecosystem. Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70B. Your first encounter with them is a model too large to run on consumer hardware.
Meanwhile, labs like Tencent and Alibaba Qwen cover the full spectrum from sub-1B upward. Two strategies, two different bets on where open-weight value accumulates.
American Open Source Moved to Hardware Vendors
The U.S. isn't absent from open source—it's just playing a different game now. The two organizations publishing the most new open models this year are AMD and NVIDIA. Each released more than 200 new model repositories.
This is the new logic of open weights for hardware vendors: a model optimized for your hardware and freely available is proof the hardware works. It's demand generation for chips, not model licensing.
Google and Meta now rank well below NVIDIA in new model releases, despite being the companies that defined open model publishing in previous years. Meta's move toward closed flagship models makes this shift even sharper.
Open source has moved from model labs to hardware and infrastructure companies.
When you include smaller models, embeddings, and legacy vision/speech models—where Google, Microsoft, IBM Granite, and OpenAI still generate hundreds of millions of downloads annually—U.S. participation is still growing. But the center of gravity has shifted.
Attention ≠ Adoption (And It Matters)
Here's a split that clarifies everything: Hugging Face took the top 25 model repositories by downloads accumulated in 2026 and the top 25 by likes. Exactly one repository appears in both lists.
Not a single model published in 2026 reaches the download top 25, while thirteen of the twenty-five date from 2022. all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against 5,156 likes. Kimi-K3 was pulled about 60 times per like.
Likes say a release matters—they go to frontier models in the weeks after they ship. Downloads say something is wired into a pipeline that runs on a schedule, accruing to small, stable models over years.
The Publisher Split
Chinese frontier labs are the only accounts on the Hub where the heavy band carries the volume. Effectively all of MiniMax's 2026 downloads are of models above 70B, along with 88% of Moonshot's and 55% of DeepSeek's.
No large American account looks like this. Google, Microsoft, and IBM Granite record essentially none of their 2026 downloads above 70B. NVIDIA and Meta sit at 14% and 9%.
The total download difference is stark: Moonshot's frontier-only portfolio recorded 37M downloads over the year. Qwen's broader release strategy across model sizes reached 2,045M—about 55 times more.
Open Weights Shift Where Value Accumulates
If frontier models were a licensing business, you'd expect the biggest releases to carry the tightest terms. The data shows the opposite.
Of 178 Chinese releases above 20B parameters in 2026, 59% carry Apache 2.0 and 22% carry MIT. Exactly none carry a non-commercial restriction. DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT.
On the American side of the same size band, 29% is Apache or MIT, 41% sits under custom terms, and 30% declares nothing at all.
Whatever these Chinese releases are for, it's not license revenue. The weights are given away on the most permissive terms available. The return has to come from somewhere else: API and cloud business, hardware positioning, or ecosystem control itself.
Qwen Has Become the Community's Base Model
Qwen-based models now account for 151,448 derivatives on the Hub. That's 2.6× Meta's total footprint and 4.7× the Llama repositories specifically. Google follows with 82,506 derivatives.
Qwen derivatives increased at roughly 180–210 new repositories per day throughout the first seven months of 2026. This isn't driven by individual launches—Qwen has become part of the default workflow for developers deciding what to fine-tune and deploy.
Three factors made this happen:
- Consistency: Regular release cadence, continuous updates rather than occasional flagship drops
- Coverage: Models across a wide range of sizes and use cases, letting developers stay in one ecosystem
- Openness: Apache 2.0 licensing reduces friction for modification, redistribution, and commercial use
These factors reinforce each other. A broad model family attracts more developers; more developers create more derivatives; those derivatives make the ecosystem more attractive to future users.
And this position was built largely by the community. The 151,448 derivatives represent downstream work created by other developers, not releases produced by Qwen itself. Even among the 28,531 GGUF conversions of Qwen models on the Hub, Qwen published only 54.
Small Models Remain the Practical Layer
Among models that declare a parameter count, those under 1B take 83% of all-time downloads. Everything above 100B takes 1%. Restricting to downloads accumulated in 2026 changes nothing: 3% of the volume goes to models above 70B.
Small models are the only ones that run on the hardware most developers actually have. So how does a trillion-parameter model reach anyone at all?
Through llama.cpp
In February the ggml team joined Hugging Face, with the project remaining fully open-source and community-governed. What changed is that the most important project in local inference now has durable resources behind it.
The ceiling moved with llama.cpp. The July snapshot carries GGUF builds of DeepSeek-V4-Flash at roughly 284B parameters and Kimi-K3 at roughly 2.8 trillion.
Local inference used to mean an 8B model on a laptop. It now means a trillion-parameter mixture-of-experts spread across a few consumer machines. That's the alternative route the frontier didn't have a year ago, and the reason a frontier-first release strategy is viable at all.
What This Means Going Forward
The open-weight landscape has fundamentally restructured in 2026:
- The frontier moved: China dominates the top end of parameter scale and permissive licensing
- U.S. labs pivoted: Hardware vendors replaced model labs as the primary open-source publishers
- Qwen won the ecosystem: It's the family developers standardize on, not Llama
- Quantization unlocked scale:
llama.cppmade trillion-parameter models accessible to prosumers - Small models still dominate usage: 83% of downloads go to models under 1B parameters
The gap between what gets attention (frontier models, likes) and what gets used (small models, downloads) has never been wider. If you're building, you need to know which metric actually matters for your use case.
And if you're tracking where the open ecosystem is heading, the answer is clear: it's being built on Chinese base models, quantized by community tools, deployed by hardware vendors optimizing for their chips, and running overwhelmingly as small models in production.
That's the state of open models in summer 2026. The frontier shifted, and the implications are still unfolding.