There is a quiet fact buried in Hugging Face's State of Open Models report, published on 14 August 2026, that matters more than any benchmark leaderboard.
Qwen, Alibaba's open model family, recorded 2.045 billion downloads on Hugging Face in 2026. Over the same period, Google's open models logged about 418 million and Meta's about 227 million. On the format most associated with actually running a model on your own machine — GGUF, consumed by llama.cpp, Ollama and LM Studio — Qwen sees roughly 39.6 million pulls a month, against Gemma's 20.8 million and Llama's 7.5 million.
Alibaba's own figure, using a different counting basis, is more than 3 billion downloads in six months. Both can be accurate. Neither changes who is first.
Key takeaways
- Distribution: Qwen logged 2.045 billion Hugging Face downloads in 2026 — about 4.9 times Google's open models and roughly nine times Meta's.
- Derivatives: 151,448 Qwen-based derivative models on the Hub, versus Google's 82,506 and Meta's roughly 58,000. Alibaba puts the global derivative count above 300,000.
- Licensing: among Chinese model releases above 20 billion parameters in 2026, about 59% use Apache 2.0 and 22% use MIT. Among comparable US releases, roughly 29% are Apache or MIT, 41% carry custom terms, and 30% declare no licence at all.
- Catalogue: Alibaba has open-sourced more than 460 models spanning coding, vision, speech, reasoning and on-device sizes.
- What developers actually download: 83% of all-time downloads are models under 1 billion parameters; models over 100 billion account for about 1%.
Why licence beats benchmark
A benchmark tells you which model is best today. A licence tells you whether you are allowed to build a business on it.
The permissive-licensing asymmetry in the Hugging Face data is the single most decision-relevant fact in the report. A company in Nairobi, Jakarta or São Paulo that wants to fine-tune a model for local languages, run it in-country for data-residency reasons, and ship it commercially can read the Apache 2.0 text and proceed. A custom licence with field-of-use restrictions requires a legal conversation most small teams cannot afford.
Hugging Face's own conclusion is the one worth quoting: Qwen has become part of the "default workflow" for developers deciding what models to fine-tune and deploy.
Defaults are the most durable moat in software. Downloads can spike on curiosity. A workflow embeds a model family into datasets, training scripts, quantisation tools and deployment templates — and once it is there, a competitor must displace an entire toolchain, not just outscore a benchmark.
Small models are the real story
The parameter race makes headlines; the download data tells a different story. Models under 1 billion parameters account for 83% of all-time downloads. Only about 3% of 2026 download volume came from anything above 70 billion parameters.
This is not a claim that small models are more capable. It is a claim about what the open ecosystem is for. A team building offline document retrieval, an in-vehicle assistant, a factory inspection system or a game NPC does not need a frontier model. It needs something that fits in available VRAM, runs at acceptable latency, and is legally uncomplicated.
Chinese labs ship across that entire range — from hundreds of billions of parameters down to phone-sized quantised variants — under one permissive licence. That breadth, not any single flagship, is what produced 151,000 derivatives.
The caveats, stated plainly
Download counts are not deployment counts. Hugging Face itself notes that its numbers do not capture API usage, private deployments or models distributed elsewhere, and Alibaba runs its own distribution platform, ModelScope. Alibaba's 3 billion figure and Hugging Face's 2.045 billion use different windows and filters; name your source and do not average them.
There is also a live policy question. Chinese policymakers have weighed restrictions on overseas access to the country's strongest models. The reach that produced these numbers is a strategic asset, and it is not guaranteed to remain unrestricted in both directions.
For now, the developer default has moved. It moved because of supply, licensing and tooling — and it did not require anyone to win a debate.
Sources: Hugging Face, State of Open Models (14 August 2026); Alibaba disclosures; Bloomberg and Tech in Asia reporting, August 2026.