State of Open Models: Summer 2026 Observations
2026-08-20 · Hugging Face
State of Open Models: Summer 2026 Observations
Published: August 14, 2026
Rapid Growth with Extreme Concentration
The Hugging Face Hub continued its explosive growth in the first eight months of 2026. Public model repositories increased from 2.43 million to 2.96 million. Datasets grew from 711,000 to 1 million, while Spaces expanded from 1.00 million to 1.44 million.
Despite the impressive headline numbers, the underlying distribution remains extremely skewed. Roughly 85.6% of models have fewer than 200 lifetime downloads, and just 1.5% of repositories account for 99.2% of all downloads. All observations in this report occur within this highly concentrated usage pattern.
1. The Frontier Is Moving Fast
The traditional progression path — releasing smaller models first and gradually scaling up — has been abandoned by several Chinese laboratories in 2026.
In almost every month, the largest and most performant open model released by a Chinese lab exceeded the largest model released by any American lab. Chinese monthly ceilings ranged between 754B and 2.78 trillion parameters. In contrast, U.S. models stayed under 130B in five of seven months, with notable exceptions being NVIDIA’s Nemotron 3 Ultra (561B) in May and June, and Inkling from Thinking Machines Lab.
Chinese labs have split into two distinct strategic camps:
- Frontier-only: Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70B. A developer’s first encounter with these organizations is typically a model too large to run locally.
- Full-spectrum: Tencent and Alibaba’s Qwen series cover the entire range from under 1B parameters upward.
Two developments enabled the frontier-only approach. First, building at massive scale is no longer a unique differentiator — Xiaomi, Ant Group, and Meituan all crossed the trillion-parameter mark this year. Second, the community’s quantization ecosystem now makes large models runnable on consumer hardware within days of release.
Consequently, model size has become a statement of strategic intent rather than mere technical capability. A frontier-only portfolio bets on benchmark leadership and API demand. A full-spectrum portfolio aims to become the standardized model family for developers.
The United States is not absent from open source. The two organizations publishing the most new open model repositories in 2026 are hardware vendors AMD and NVIDIA, each releasing over 200 new repositories. LiquidAI ranks third with approximately 100. Hardware companies have recognized that open, hardware-optimized models are among the most effective ways to demonstrate and sell their chips.
When smaller models and embeddings are included, U.S. organizations including Google, Microsoft, IBM Granite, and older OpenAI vision and speech models continue to generate hundreds of millions of annual downloads.
At frontier scale, several major U.S. releases above 100B parameters this year were built on top of Chinese models or leveraged artifacts from Chinese labs (e.g., Thinking Machines’ Inkling at 952B). Notable original American large models include NVIDIA’s Nemotron 3 Ultra (561B), Nemotron 3 Super (124B), and Arcee AI’s Trinity-Large (399B). AMD has contributed extensive conversion and optimization work, enabling trillion-parameter models to run efficiently on U.S. hardware.
Meanwhile, Chinese open models are increasingly optimized for domestic Chinese chips, creating a mirrored hardware-model ecosystem competition.
2. Attention Does Not Equal Adoption
When comparing the top 25 model repositories by downloads accumulated during 2026 against the top 25 by likes, only one repository appears on both lists.
The analysis uses downloads within the 2026 window rather than lifetime totals, removing the advantage of simply having existed longer. No model published in 2026 reached the top 25 by downloads, while thirteen of the top 25 were originally released in 2022.
The patterns are stark: all-MiniLM-L6-v2 was downloaded 1.55 billion times in seven months with only 5,156 likes. Some frontier models receive roughly 60 downloads per like.
These metrics measure different behaviors. Likes signal that a release matters and typically go to frontier models shortly after launch. Downloads indicate that a model has been integrated into scheduled production pipelines and accrue over years to small, stable, reliable models.
This split is also visible at the publisher level. Chinese frontier labs are the only accounts where the majority of 2026 downloads come from large models: nearly 100% for MiniMax, 88% for Moonshot, 55% for DeepSeek, and 39% for Z.ai. No major U.S. publisher shows this pattern — Google, Microsoft, and IBM Granite recorded essentially zero downloads above 70B in 2026, while NVIDIA and Meta recorded only 14% and 9% respectively.
The absolute gap is even more pronounced. Moonshot’s frontier-only portfolio recorded approximately 37 million downloads for the year, while Qwen’s broad release strategy across all model sizes reached 2.045 billion downloads across its repositories.