Open artificial intelligence is not advancing in just one direction. While some labs compete with models containing hundreds of billions or even trillions of parameters, daily adoption remains concentrated around small, stable models that are easy to run. So, what is really happening behind the numbers?
The State of Open Models: Summer 2026 Observations report, published by Hugging Face, analyzes activity on the Hub during the first seven months of 2026. Its data shows an ecosystem that is larger and more competitive, but also more dependent on community tools such as llama.cpp, the GGUF formats, and coding agents.
The open ecosystem is growing, but unevenly
Between late 2025 and July 2026, public model repositories on Hugging Face grew from 2.43 million to 2.96 million. Datasets increased from 711,000 to 1 million, while Spaces—the applications and demos hosted on the platform—grew from 1 million to 1.44 million.
However, growth does not mean that every project has the same visibility. A total of 85.6% of models have recorded fewer than 200 cumulative downloads throughout their entire existence. At the same time, just 1.5% of repositories account for 99.2% of all downloads.
The Hub is enormous, but its activity is concentrated in a small fraction of models and tools.
This difference matters. A model can receive thousands of “likes” because it excites the community and still have few downloads. Another can be downloaded millions of times without attracting much attention because it is already part of an application, workflow, or software library.
China dominates the open-model frontier
During almost every month of 2026, the largest and most powerful open model published by a Chinese lab surpassed the largest original model introduced by a US company. China’s monthly ceiling ranged from 754 billion to 2.78 trillion parameters.
In the United States, that ceiling remained below 130 billion parameters in five of the seven months analyzed. The exceptions were Nemotron 3 Ultra, from NVIDIA, with 561 billion parameters, and Inkling, from Thinking Machines Lab.
This comparison does not mean that all Chinese models are better at every task. A model’s size alone does not determine its quality, speed, cost, or usefulness. But it does show a shift in publishing strategy: several Chinese labs jumped directly to the frontier instead of first building a ladder of smaller models.
Moonshot, MiniMax, Xiaomi, and Z.ai mainly publish models with more than 70 billion parameters. For an independent developer, that means their first contact with these families may be a model that is impossible to run on a personal computer without quantization, multiple GPUs, or external services.
Alibaba, with Qwen, and Tencent are following a different strategy. They publish models across a wide range, from fewer than 1 billion parameters to frontier-scale sizes. This makes the family a more flexible option for anyone who needs anything from a local assistant to a high-performance solution.
The United States is changing its role in open AI
The United States has not disappeared from the open ecosystem. In fact, AMD and NVIDIA were the organizations that published the most new models in 2026, with more than 200 repositories each. LiquidAI ranked third, with around 100.
The reason is fairly practical: open models also help sell hardware. If a company publishes a model optimized for its chips and anyone can try it, it is directly demonstrating that its infrastructure works.
When small models, embeddings, and older vision or speech models are included, Google, Microsoft, IBM Granite, and OpenAI still generate hundreds of millions of downloads each year. But the center of gravity has shifted. Google and Meta publish fewer new models than NVIDIA, and Meta is moving toward a more closed strategy for its main models.
The difference is even clearer above 100 billion parameters. Most US launches in that range are not original models, but conversions, optimizations, or adaptations of Chinese models. Exceptions include Inkling, Nemotron 3 Ultra, Nemotron 3 Super, and Trinity-Large, from Arcee AI.
AMD, for example, contributed many conversions but no original model of that size. Its work remains relevant because it makes it possible to run Chinese trillion-parameter models more efficiently on US hardware. It is a different role: distribution and optimization, not creation of the base model.
“Likes” show enthusiasm; downloads show dependence
Hugging Face compared the 25 repositories with the most cumulative downloads during 2026 with the 25 that received the most “likes.” Only one appeared on both lists.
No model published in 2026 made it into the top 25 for downloads. Thirteen of those 25 repositories date back to 2022. all-MiniLM-L6-v2, a small model used for search and semantic similarity tasks, accumulated 1.55 billion downloads in seven months, even though it received only 5,156 “likes.”
The takeaway is simple:
- A “like” indicates community interest, relevance, or enthusiasm.
- A download usually indicates that the model is connected to an application or repetitive process.
- New models attract attention quickly.
- Small, stable models accumulate usage over the years.
Confusing these two metrics can lead to the wrong conclusions. The model generating the most conversation is not necessarily the one doing the most real work.
Qwen is winning as an ecosystem, not just as a model
Qwen’s strategy stands out because it does not depend solely on launching a huge model. Its models appear in 151,448 derivatives on Hugging Face—2.6 times Meta’s total footprint and 4.7 times the number of repositories derived from Llama.
Google ranks second, with 82,506 derivatives. One notable example is Unsloth, a community account that publishes quantized versions prepared for fine-tuning. Many of these models further expand the Qwen ecosystem.
During the first seven months of 2026, Qwen derivatives grew at an approximate rate of 180 to 210 new repositories per day. This suggests that Qwen is becoming a common foundation for deciding which model to fine-tune, adapt, and deploy.
Why does this strategy work? Because of a combination of factors:
- Consistency: it maintains a regular release schedule.
- Coverage: it offers models for different sizes, tasks, and budgets.
- Open licenses: Apache 2.0 makes modification, redistribution, and commercial use easier.
- Network effect: each derivative increases the ecosystem’s value for the next developers.
The figure of 151,448 derivatives does not represent models created directly by Qwen. It is work produced by the community. Even among the 28,531 conversions to GGUF, Qwen officially published only 54.
Small models still dominate real-world use
Models with fewer than 1 billion parameters account for 83% of historical downloads among repositories that disclose their parameter count. Models with more than 100 billion account for barely 1%.
When looking only at 2026 downloads, the conclusion changes very little: just 3% of the volume corresponds to models with more than 70 billion parameters.
There is a physical explanation for this, not just a commercial one. Most developers do not have access to large GPU clusters. They need models that can run on a laptop, an affordable server, or a moderate cloud instance.
So, how does a 2.8 trillion-parameter model reach a personal computer? The answer is llama.cpp, the local inference tool that makes it possible to run quantized models on consumer hardware and distribute large models across multiple machines.
In July, GGUF versions of DeepSeek-V4-Flash, with around 284 billion parameters, and Kimi-K3, with approximately 2.8 trillion, appeared. Local inference no longer means only running an 8-billion-parameter model on a laptop. It can also mean distributing a mixture-of-experts model across several computers.
The addition of the ggml team to Hugging Face in February strengthened this trend. The project remains open and community-governed, but it now has more lasting resources behind it.
The execution layer is growing faster than the core
Repositories that report using the gguf library grew 464% during the period analyzed. Those related to lerobot increased 194%, while Apple mlx repositories grew 148%.
By comparison, transformers and peft grew 16%, while diffusers advanced 21%. The conclusion is revealing: the layer that determines where and how a model runs is growing three to seven times faster than the modeling core.
For labs, this presents a concrete opportunity. Publishing an official GGUF conversion, documenting quantization options, and signing artifacts requires far less effort than training a new model. It would also bring the version tested by its creators closer to the version the community actually uses.
Another option would be to collaborate with projects such as Unsloth and other conversion maintainers. A model does not end when its weights are published. It also needs formats, tools, documentation, and compatibility with the hardware people have available.
Agents are already starting to use Hugging Face
In July, Hugging Face published a dataset on agent usage. This record identifies the agent/<name> token that coding agents send when using huggingface_hub or the hf tool to search for models, upload data, run Jobs, or create Spaces.
Claude Code led July with 44.4% of identified traffic. But the most interesting detail appears when looking at the evolution over time: it held 67.8% in April and barely 6.4% in May. Codex, meanwhile, rose from 10.4% to 20.8%.
There is no permanent leader yet. A new launch or a change in default configuration can move a huge share of traffic in just one month.
In addition, nearly a quarter of agent-labeled traffic in July came from tools that still did not have a registered name. In May, that proportion reached 59.8%. More than a dozen new identifiers appeared between April and July.
Hugging Face has also begun adapting its platform for non-human readers. It published machine-readable Markdown, added agent traces, introduced an agents.md endpoint in every Gradio Space, and added hf_fs to its MCP server to expose repositories, storage, documentation, and articles through a single interface.
The trend points to an important shift: agents are not only consulting the Hub. They can also search for models, run code, create applications, and complete tasks within the ecosystem.
The security case that changed the conversation
In July, Hugging Face documented what it describes as one of the first known cases of an autonomous agent carrying out a sustained intrusion on its own initiative.
The episode revealed a paradox. When the team tried to use closed frontier models to analyze the captured code, their safety systems rejected the request. The analysis ultimately took place with a quantized open model, GLM-5.2, running on its own infrastructure.
The case does not prove that open models are automatically safer or more dangerous. It does show that control over infrastructure can be decisive in response, auditing, and security tasks. A local system does not depend on the same access policy as a commercial API, although it still requires appropriate human and technical safeguards.
Licenses reveal a bet on the ecosystem
Of 178 Chinese models published in 2026 with more than 20 billion parameters, 59% use Apache 2.0 and 22% use MIT. None includes an explicit noncommercial-use restriction.
DeepSeek and Z.ai even published models ranging from 700 billion to 1.65 trillion parameters under MIT. In the same range, US launches were much more restrictive or ambiguous: only 29% used Apache or MIT, while 41% used custom terms and 30% did not declare a license.
If a lab releases giant models under permissive licenses, it is probably not seeking direct licensing revenue. The bet may be on APIs, the cloud, hardware, the company’s valuation, or the position it occupies within the ecosystem.
The question is no longer only who has the most powerful model. It also matters who can get thousands of developers to adapt it, integrate it, and turn it into everyday infrastructure.
What these data points do allow us to conclude
Hugging Face’s report does not measure the entire artificial intelligence industry. Its downloads do not necessarily include API usage, private deployments, or models distributed outside the Hub. “Likes” measure attention, not quality, and derivatives show development activity, not revenue.
Even so, the signals are consistent. China is pushing hard at the open frontier. The United States retains an important role, especially through hardware and infrastructure. Qwen is becoming a development foundation. Small models still dominate everyday use, and local inference tools are growing quickly.
The competition is not limited to training models either. It is being fought through licenses, conversions, formats, agents, chips, and the communities that build on every release.
In artificial intelligence, a few months can completely change the map. But the model that generates the most excitement today will not always be the one supporting tomorrow’s applications. In the long run, the advantage may belong to the family that developers can run, adapt, and turn into part of their daily work.
Original source
https://huggingface.co/blog/state-of-open-models-summer-2026