Glonce

Hugging Face: Qwen becomes the base model of open source AI

modelsThe Glonce editorsHugging Face🔥78Glonce rating78Impact72Novelty70Relevance95Surprise65Players85Credibility85
Hugging Face: Qwen becomes the base model of open source AI

Hugging Face published a Hub-data review covering the first seven months of 2026. Public model repositories on the Hub grew from 2.43 to 2.96 million, datasets from 711,000 to 1 million, and Spaces from 1.00 to 1.44 million. The activity underneath stays concentrated: about 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.

At the frontier, China's monthly ceiling for the largest open model released ran between 754B and 2.78 trillion parameters, while America's own ceiling stayed under 130B in five of seven months. The exceptions were NVIDIA's Nemotron 3 Ultra at 561B in May and June and Thinking Machines' Inkling at 952B. Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70B, so a developer's first encounter with them is a model too large to run locally; Tencent and Alibaba's Qwen instead cover the whole range from under 1B upward. Xiaomi and Meituan both crossed a trillion parameters this year despite not being known names in open weights twelve months ago. Most US releases above 100B parameters this year are conversions of Chinese models rather than new ones; the few original American frontier models are Inkling (952B), Nemotron 3 Ultra (561B), Nemotron 3 Super (124B) and Arcee AI's Trinity-Large (399B). AMD and NVIDIA are this year's two most prolific publishers of new open model repositories, each releasing more than 200, with LiquidAI third at around 100; Google and Meta now rank below them, a reversal from the years they defined open model publishing, reinforced by Meta's move toward closed flagship models.

On usage, the top 25 repositories by 2026 downloads and the top 25 by likes share exactly one repository. No model published in 2026 reaches the download top 25, while thirteen of the twenty-five date from 2022. all-MiniLM-L6-v2 was downloaded 1.55 billion times in seven months against 5,156 likes; Kimi-K3 was downloaded about 60 times per like it received. Chinese frontier labs concentrate their downloads in large models: effectively all of MiniMax's 2026 downloads, 88% of Moonshot's, 55% of DeepSeek's and 39% of Z.ai's come from models above 70B, against roughly none for Google, Microsoft and IBM Granite and only 14% for NVIDIA and 9% for Meta. Moonshot's frontier-only strategy brought 37 million downloads over the year; Qwen's broader strategy across sizes brought 2,045 million (2,061 million including repositories without a declared parameter count), about 55 times more.

Chinese frontier releases are also licensed loosely. Of 178 Chinese releases above 20B parameters this year, 59% carry Apache 2.0 and 22% carry MIT, with none under a non-commercial restriction; DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT. American releases in the same size band are more restrictive: 29% Apache or MIT, 41% under custom terms, 30% with no license declared at all. Hugging Face argues the payoff for the Chinese labs is not licence revenue but comes from elsewhere: API and cloud business, hardware and platform positioning, or ecosystem standing, pointing to Z.ai and Kimi's valuations as evidence the open strategy is working, while expecting clearer monetization paths to emerge over time.

Qwen has become the largest foundation in the open model ecosystem by derivative count: 151,448 Qwen-based repositories on the Hub, 2.6 times Meta's total footprint and 4.7 times the Llama repositories specifically. Google follows with 82,506 derivatives, and Unsloth, a community account publishing quantized and fine-tuning-ready builds, ranks third. Qwen derivatives have grown at roughly 180 to 210 new repositories a day through the first seven months of 2026. Hugging Face credits Qwen's steady release cadence, its coverage across model sizes, and its Apache 2.0 licensing. Most of that footprint is downstream: of 28,531 GGUF conversions of Qwen models on the Hub, Qwen itself published only 54.

Small models remain where actual usage sits: among models with a declared parameter count, those under 1B take 83% of all-time downloads and everything above 100B takes 1%, a split that barely moves when restricted to 2026 downloads alone (3% above 70B). The route by which huge models reach ordinary hardware is llama.cpp: the ggml team behind it joined Hugging Face in February 2026 while staying fully open source and community governed. The July 2026 snapshot carries GGUF builds of DeepSeek-V4-Flash at roughly 284B parameters and Kimi-K3 at roughly 2.8 trillion, letting mixture-of-experts models that size run across a few consumer machines. Qwen leads monthly GGUF downloads at 39.6 million, nearly twice Gemma's 20.8 million and more than five times Llama's 7.5 million, even though Llama-derived GGUF repositories slightly outnumber Qwen's, so the gap is not a supply problem. Overall repository count grew 21.5% over the seven months, but the infrastructure layer grew faster: repositories declaring the gguf library rose 464%, lerobot 194% and Apple's mlx 148%, against 16% for transformers and peft and 21% for diffusers. Hugging Face notes that the labs behind the ten largest model families rarely publish official GGUF conversions themselves, leaving that work to the community despite the limited extra effort it would take.

Key facts

  • China's monthly ceiling for the largest open model released ran from 754B to 2.78 trillion parameters in 2026, while the US ceiling stayed under 130B in five of seven months.
  • Qwen-based derivatives on the Hub reached 151,448, 2.6 times Meta's total footprint and 4.7 times the Llama repositories specifically, making Qwen the ecosystem's largest base model by downstream use.
  • Only one repository appears on both the top-25-by-downloads and top-25-by-likes lists for 2026, and no model published in 2026 reaches the download top 25.
  • Of 178 Chinese releases above 20B parameters, 59% carry Apache 2.0 and 22% MIT with zero non-commercial restrictions, against 29% Apache or MIT among comparable US releases.
  • AMD and NVIDIA each published more than 200 new open model repositories this year, more than any AI lab, while Google and Meta's release counts fell behind them.

Why it matters

The report describes a changed shape for open source AI in 2026. Frontier-scale releases are now led by Chinese labs, several of which skipped the traditional path of starting small and scaling up; their monthly ceiling for the largest released model stayed above America's for most of the year. In the US, the two organizations publishing the most new open models are the hardware makers AMD and NVIDIA, not the traditional model labs, while Google and Meta have fallen behind them and Meta has moved its flagship models to closed release. Separately, Qwen has become the default base layer developers build on, with far more downstream derivatives than any other family, including Meta's Llama.

Who it affects

Developers picking a base model to fine-tune or deploy, since Qwen's release cadence, size coverage and licensing have made it the most-built-upon option on the Hub. Hardware vendors AMD and NVIDIA, who are using open model releases to demonstrate that their chips work. US labs positioning themselves against Chinese frontier releases, particularly since most US releases above 100B parameters this year are conversions of Chinese models rather than original ones. Anyone running large models locally, who depends on the llama.cpp and GGUF ecosystem to make trillion-parameter models usable on consumer hardware.

How to use it

Most large Chinese model releases carry permissive licenses: 59% Apache 2.0 and 22% MIT among 178 Chinese releases above 20B parameters this year, with no non-commercial restrictions, including DeepSeek and Z.ai models between 700 billion and 1.65 trillion parameters under plain MIT. For fine-tuning and deployment, Qwen's broad size range and steady updates make it the practical default base, while for actual local deployment, models under 1B parameters still account for 83% of all-time downloads. Trillion-parameter models such as Kimi-K3 (about 2.8 trillion parameters) are usable locally through GGUF quantization builds via llama.cpp rather than in their original form.

How solid is it

The figures come directly from Hugging Face's own Hub telemetry: repository, download and like counts the company tracks across its own platform, covering a defined seven-month window in 2026. The comparison between downloads and likes explicitly controls for a model's age to avoid crediting older repositories for having existed longer. The post carries no named author or byline. Some of its framing, such as why Google and Meta's new-model counts fell behind NVIDIA and AMD, and its expectation that monetization paths will become clearer, is the report's own interpretation rather than a measured figure.

Risks and caveats

The post does not explain why Google's and Meta's new-model-release counts dropped behind NVIDIA's beyond noting Meta's shift toward closed flagship models. It gives no timeline for when clearer open source monetization might appear, and does not name which ten model families it means by the largest families that rarely publish official GGUF conversions. All figures are drawn from Hugging Face's own Hub, so they measure activity on that platform specifically, not usage through APIs or other distribution channels. Several numbers are stated as approximations, such as the 754B to 2.78 trillion parameter range for China's monthly ceiling and the roughly 284B and 2.8 trillion parameter counts for the July GGUF builds.

“Whatever these releases are for, it is not licence revenue.”

— Hugging Face, "State of Open Models: Summer 2026 Observations"