Glonce

Hugging Face finds Qwen has become open source AI's base model

modelsThe Glonce editorsHugging Face🔥77Glonce rating77Impact70Novelty65Relevance100Surprise60Players85Credibility90
Hugging Face finds Qwen has become open source AI's base model

Hugging Face's own mid-year analysis of its model Hub tracks growth and usage patterns across the first seven months of 2026. Public model repositories grew from 2.43 million to 2.96 million, datasets from 711,000 to 1 million, and Spaces from 1.00 million to 1.44 million. Usage stays extremely concentrated: about 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.

At the frontier, the report finds that several Chinese labs skipped the traditional progression of shipping small models before working up to flagship size. In almost every month of 2026, the largest open model released by a Chinese lab outsized any American release that month. China's monthly ceiling for the largest released model ranged from 754 billion to 2.78 trillion parameters, while U.S. labs stayed under 130 billion parameters in five of the seven months, the exceptions being NVIDIA's Nemotron 3 Ultra at 561 billion parameters in May and June, and Thinking Machines Lab's Inkling. Hugging Face splits Chinese publishers into two camps: Moonshot, MiniMax, Xiaomi and Z.ai release almost nothing below 70 billion parameters, so a developer's first encounter with them is typically a model too large to run locally, while Tencent and Alibaba's Qwen cover the entire range down to under 1 billion parameters. The report attributes the frontier-only strategy partly to community quantization work, which now makes a large model runnable within days of release even if the lab itself never ships a small version.

The U.S. is not absent from open weights, but its presence looks different. AMD and NVIDIA are described as the two organizations publishing the most new open model repositories in 2026, each releasing more than 200, far ahead of the rest of the field; LiquidAI ranks third with around 100. Hugging Face frames this as hardware vendors using open models to sell chips. Major original American frontier releases named in the report include NVIDIA's Nemotron 3 Ultra (561B) and Nemotron 3 Super (124B), and Arcee AI's Trinity-Large (399B); some other U.S. releases above 100 billion parameters, such as Thinking Machines' Inkling (952B), are built on top of Chinese models or use artifacts from Chinese labs. AMD is credited with many of the conversions that let trillion-parameter models run efficiently on U.S. hardware, work the report calls a distribution and optimization layer rather than original model development. Chinese labs, meanwhile, are increasingly optimizing their own open models for domestic chips, the same dynamic playing out in reverse.

A separate section distinguishes attention from actual use. Comparing the top 25 model repositories by downloads accumulated in 2026 against the top 25 by likes, Hugging Face finds exactly one repository appears on both lists; no model published in 2026 reaches the download top 25, while thirteen of the twenty-five download leaders date from 2022. The embedding model all-MiniLM-L6-v2 was downloaded 1.55 billion times in seven months against 5,156 likes, while Kimi K3 was downloaded about 60 times for every like it received. The same split shows up at the publisher level: essentially all of MiniMax's 2026 downloads, 88% of Moonshot's, 55% of DeepSeek's and 39% of Z.ai's come from models above 70 billion parameters, while Google, Microsoft and IBM Granite record almost none of their downloads above that threshold, and NVIDIA and Meta only 14% and 9%. In total downloads, Moonshot's frontier-only portfolio brought in 37 million downloads over the year, versus 2,045 million for Qwen across repositories with a declared parameter count (2,061 million including all repositories), about 55 times more.

On licensing, the report finds that of 178 Chinese model releases above 20 billion parameters in 2026, 59% carry Apache 2.0 and 22% carry MIT, with most others under non-commercial restrictions; DeepSeek and Z.ai ship models between 700 billion and 1.65 trillion parameters under plain MIT, making their largest releases about as permissively licensed as their smallest. On the U.S. side of the same size band, only 29% is Apache or MIT, 41% sits under custom terms and 30% declares no license at all. Hugging Face notes a recent shift, however: Kimi K3 and Qwen3.8 have begun adding non-commercial restrictions and revenue-share requirements to their licenses, and the report expects the industry to move toward clearer monetization paths going forward.

The report's central claim about Qwen is that it has become the community's base model. Qwen-based derivatives on the Hub now number 151,448, which Hugging Face puts at 2.6 times Meta's total derivative footprint and 4.7 times the Llama repositories specifically; Google is second with 82,506 derivatives, and the community account Unsloth, which publishes quantized and fine-tuning-ready builds, is third, with much of its output itself building on Qwen. New Qwen derivatives have appeared at roughly 180 to 210 repositories per day through the first seven months of 2026. Hugging Face credits three factors: a steady Qwen release cadence rather than occasional flagship drops, coverage across a wide range of model sizes, and Apache 2.0 licensing that lowers friction for reuse. It stresses that the 151,448 figure is downstream community work, not Qwen's own output: of 28,531 GGUF conversions of Qwen models on the Hub, Qwen itself published only 54.

The report's other repeated finding is that small models remain the practical layer. Among models that declare a parameter count, those under 1 billion parameters take 83% of all-time downloads, while everything above 100 billion takes 1%; restricting to 2026-only downloads, models above 70 billion still take just 3%. Hugging Face credits llama.cpp for extending that ceiling: after the ggml team behind the project joined Hugging Face in February 2026, while llama.cpp stayed fully open-source and community-governed, the July 2026 snapshot already carried GGUF builds of DeepSeek-V4-Flash at roughly 284 billion parameters and Kimi K3 at roughly 2.8 trillion, letting trillion-parameter mixture-of-experts models run spread across a few consumer machines. Qwen leads this local-inference layer too, at 39.6 million GGUF downloads a month against Gemma's 20.8 million and Llama's 7.5 million; the report says the Llama gap is not a supply problem, since Llama-derived GGUF repositories slightly outnumber Qwen's. Infrastructure around local inference is growing faster than the modelling core: over the seven months, model repositories overall grew 21.5%, while 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 ten largest model families still publish very few official GGUF conversions themselves, even though those conversions are what most developers running models locally actually use.

Key facts

  • China's monthly ceiling for the largest released open model ran from 754 billion to 2.78 trillion parameters in 2026, while U.S. labs stayed under 130 billion parameters 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, with Google second at 82,506.
  • Qwen's total 2026 downloads reached 2,045 million (2,061 million including all repositories), about 55 times Moonshot's 37 million, despite Moonshot's larger, frontier-only model portfolio.
  • Among models with a declared parameter count, those under 1 billion take 83% of all-time downloads while everything above 100 billion takes just 1%.
  • Only one repository appears in both the top 25 by 2026 downloads and the top 25 by likes; thirteen of the download leaders date from 2022, none from 2026.

Why it matters

The report matters because it replaces impressions of the open model race with the Hub's own usage data, and the picture it draws cuts against the usual framing. Chinese labs are not catching up to the frontier gradually; several skipped straight to trillion-parameter releases, while the largest American open releases mostly stayed under 130 billion parameters. At the same time, the model getting the most attention and the model actually being built on are not the same thing: Qwen, not any single flagship release, has become the base most of the ecosystem's derivative work sits on.

Who it affects

Developers picking a base model to fine-tune or deploy, since Qwen's derivative count and release cadence now make it a default choice in the way Llama once was. Hardware vendors AMD and NVIDIA, whose open model releases the report treats as a way to sell chips rather than a licensing business. Labs deciding on release strategy and licensing, as the report documents a split between frontier-only Chinese publishers and full-spectrum ones, and flags Kimi K3 and Qwen3.8 starting to add non-commercial and revenue-share terms. And anyone running models locally, since the ggml team's move to Hugging Face and the growth of GGUF conversions determine which large models are actually reachable outside a data center.

How to use it

For developers choosing a base model, Qwen's derivative count, download volume and Apache 2.0 licensing make it the low-friction default the report describes, especially for anyone planning to fine-tune or quantize. For running large models locally, GGUF builds via llama.cpp, now backed by Hugging Face after the ggml team's move, are how models like DeepSeek-V4-Flash and Kimi K3 become usable on consumer hardware despite running to hundreds of billions or trillions of parameters. Licensing should be checked model by model rather than assumed from a lab's size or country: Chinese labs license some of their largest models, such as DeepSeek and Z.ai's, under plain MIT, but Kimi K3 and Qwen3.8 have recently added non-commercial and revenue-share restrictions.

How solid is it

The data comes directly from Hugging Face about its own Hub, published on its own blog in the collective 'we' voice with no named author and no stated publication date; the methodology is described inline, including restricting download counts to the 2026 window rather than lifetime totals to control for age when comparing top lists. That first-party vantage is also a limit: the report only sees what happens on the Hub, not on other model repositories or private deployments, and Hugging Face has its own interest in a narrative where its platform sits at the center of the ecosystem. The retrieved text cuts off mid-sentence in its closing section, so whatever concrete collaboration proposal the authors make around official GGUF conversions is not available here.

Risks and caveats

The comparisons rest on specific thresholds, 70 billion and 100 billion parameters mark the dividing lines used throughout the piece, and the picture could look different at other cutoffs. Several of the fastest growth figures, such as the 464% rise in repositories declaring the gguf library or 194% for lerobot, are percentage changes off what are likely much smaller starting bases than the modelling core's 21.5% growth, so the same absolute increase looks far more dramatic as a percentage. No mechanism is offered for why Chinese labs skipped the small-to-large release progression, beyond the observations about quantization and hardware differentiation already in the report; that remains an inference, not a stated cause. And the licensing figures count stated terms, not whether a non-commercial or custom license is actually enforced in practice.