This article analyses activity on the Hugging Face Hub from January through August 2026. During that period the Hub continued to expand rapidly while the centre of gravity in open models shifted: several Chinese frontier labs repeatedly published the largest public models, and hardware and infrastructure companies strengthened their role in open releases. Small, stable models nevertheless remain the backbone of Hub usage.
Growth on the Hub — key figures
- Public models rose from 2.43 million to 2.96 million during the period.
- Datasets increased from 711,000 to 1 million.
- Spaces grew from 1.00 million to 1.44 million. The distribution underneath those totals is extreme: about 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.
1. The frontier is moving fast
In 2026 several Chinese labs abandoned the old stepwise path (release small models first, then scale up) and published very large models directly. In almost every month of 2026 the largest and highest‑performing open model from a Chinese lab exceeded anything an American lab published in the same month. China’s monthly ceiling ran between 754 billion and 2.78 trillion parameters; the U.S. ceiling stayed below 130 billion in five of seven months, with exceptions such as NVIDIA’s Nemotron 3 Ultra (561B in May and June) and Thinking Machines’ Inkling.
Labs split into two camps. Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70B, so a developer’s first encounter with them is typically a model too large to run locally. Tencent and Alibaba Qwen instead cover the whole range from under 1B upward. Two forces enabled the first camp: massive scale stopped being a unique differentiator (Xiaomi and Meituan cleared a trillion parameters this year), and community quantization/conversion tooling now makes large models runnable within days.
The United States remains active in open source, but in a changing role: AMD and NVIDIA each released more than 200 new model repositories this year, with LiquidAI around 100, showing that hardware vendors see open models as a way to demonstrate and sell chips. At frontier scale most U.S. releases above 100B this year are conversions or optimizations of Chinese models rather than original models — an important distribution and optimization role rather than primary model creation. Notable original American frontier models include Thinking Machines’ Inkling (952B), NVIDIA’s Nemotron 3 Ultra (561B) and Nemotron 3 Super (124B), and Arcee AI’s Trinity‑Large (399B).
2. Attention ≠ Adoption
Comparing the top 25 repositories by downloads (accumulated in the year) to the top 25 by likes shows almost no overlap: exactly one repository is on both lists. The analysis counts downloads within the year (not lifetime) to control for age; no model published in 2026 reaches the download top 25, while 13 of the 25 date from 2022.
Likes and downloads measure different behaviours. Likes capture attention and excitement—typically directed at frontier releases soon after shipping. Downloads capture integration into pipelines and persistent usage over time. For example, all‑MiniLM‑L6‑v2 was pulled 1.55 billion times in seven months but received only 5,156 likes; Kimi‑K3 shows a very different like‑to‑download ratio.
This split appears at the publisher level. Chinese frontier labs are the only Hub accounts whose large models carry the download volume: nearly all of MiniMax’s 2026 downloads are of models above 70B; similarly 88% of Moonshot’s, 55% of DeepSeek’s and 39% of Z.ai’s. No large American account looks like this: Google, Microsoft and IBM Granite record essentially none of their 2026 downloads above 70B; NVIDIA and Meta only 14% and 9% respectively.
A concrete contrast: Moonshot’s frontier‑only portfolio recorded 37M downloads over the year; Qwen’s broader range of releases reached 2,045M downloads (2,061M including all repositories declaring or not declaring parameters) — roughly 55× more.
3. Open weights shift where value accumulates
Licensing patterns show that many large Chinese releases are permissively licensed. Of 178 Chinese releases above 20B parameters this year, 59% use Apache 2.0 and 22% use MIT; exactly none carry a non‑commercial restriction. Chinese labs therefore license their largest models about as permissively as their smallest.
On the American side in the same size band, 29% use Apache/MIT, 41% sit under custom terms, and 30% declare no license. This suggests these large open‑weight releases are not intended to generate licensing revenue; value likely accrues via API/cloud business, hardware and platform positioning, or the ecosystem effects (traction and valuations) that permissive openness enables.
4. Qwen has become the community base model
Measured by how much the community builds on a family, Qwen has become one of the largest foundations in the open model ecosystem. Qwen‑based models now account for 151,448 derivatives on the Hub — 2.6× Meta’s total footprint and 4.7× the number of Llama repositories specifically. Google follows with 82,506 derivatives. The third largest source is Unsloth, a community account publishing quantized and fine‑tuning‑ready builds that often extend the Qwen ecosystem.
Qwen derivatives increased at roughly 180–210 new repositories per day across the first seven months of 2026, indicating adoption driven by continuous community activity rather than single flagship events. Several factors explain Qwen’s position: consistent release cadence, coverage across a wide range of sizes and use cases, and permissive Apache 2.0 licensing that lowers friction for modification and commercial use. Importantly, the 151,448 derivatives are downstream work from other developers; among 28,531 GGUF conversions of Qwen models on the Hub, Qwen itself published only 54.
5. Small models remain the practical layer
Among models that declare parameter counts, those under 1B account for 83% of all‑time downloads, while everything above 100B accounts for 1%. Restricting to downloads accumulated in 2026 produces a similar picture: only 3% of download volume goes to models above 70B. The practical reason is simple: most developers’ hardware runs smaller models.
Local inference has, however, expanded dramatically through llama.cpp and ggml. In February the ggml team joined Hugging Face while keeping the project fully open and community‑governed; the change was that durable resources now back the most important local‑inference project. The July snapshot included GGUF builds such as DeepSeek‑V4‑Flash (~284B) and Kimi‑K3 (~2.8T). The local‑inference ceiling thus moved from laptop‑scale 8B models to trillion‑parameter mixtures of experts spread across consumer machines — an alternative deployment route that did not exist a year earlier.
This route runs strongly on Qwen: Qwen GGUF downloads are ~39.6M per month, nearly twice Gemma’s 20.8M and more than five times Llama’s 7.5M. Llama‑derived GGUF repositories slightly outnumber Qwen’s, so the gap is not shelf‑space but traffic.
Model repositories grew 21.5% over the seven months, but supporting layers grew much faster: repositories declaring GGUF rose 464%, lerobot 194% and Apple‑specific mlx148% declarations grew severalfold, while transformers and peft grew ~16% and diffusers 21%. The platform layer that decides where a model can physically run (local inference formats, Apple silicon, robot control stacks) is expanding three‑to‑seven times faster than the modelling core.
Across the ten largest model families, the labs behind them publish very few official GGUF conversions, yet developers often use GGUF versions locally. Providing official conversions at release, documenting quantization decisions, and signing artifacts would require limited extra work and could narrow the gap between weights tested by model creators and those adopted by the wider community.
6. Agents are the new user
A new instrument appeared this year: the agent‑usage dataset (published in July) records agent/<name> tokens that coding agents send when they call the Hub via huggingface_hub or the hf CLI — for example when searching models, pushing datasets, running Jobs, creating Spaces. For the first time it is possible to measure agent traffic and see which harnesses it comes from.
The monthly shares are volatile. Claude Code led July with 44.4%, but it held 67.8% in April and 6.4% in May, while Codex climbed steadily from 10.4% to 20.8%. There is no stable incumbent: a single release or a changed default can shift half the traffic in a month. Also notable is the unregistered row: nearly a quarter of agent‑tagged traffic in July came from harnesses not yet named in the dataset; in May that figure was 59.8%. New entrants appear faster than registries can name them.
Tooling evolved to serve machine readers as well as human browsers: papers began serving machine‑readable Markdown in March; April added agent traces as a first‑class dataset type and an agents.md endpoint on every Gradio Space; July brought the hf_fs tool on the MCP server, exposing repositories, storage, docs and papers via a single interface and attachable sandboxes for secure execution. MCP later joined the Linux Foundation’s Agentic AI Foundation.
A darker development arrived in July: an apparent first documented case of an autonomous agent running a sustained intrusion on its own initiative against our infrastructure. Our team initially attempted analysis with closed frontier models, but safety guardrails prevented the work; the analysis was completed on a quantized open GLM‑5.2 running on our own infrastructure. We published a disclosure and a full technical timeline.
Looking forward
Compared to the spring report, geographic rebalancing of power has accelerated: U.S. open models continue to matter, but multiple Chinese frontier models now define community attention and create valuation‑and‑growth opportunities for their publishers. Likes tend to indicate short‑term excitement at the frontier; downloads, derivatives and local deployment indicate long‑term infrastructure adoption.
Tools like llama.cpp and GGUF enable deployment of very large models locally, but a broad model family and community adoption remain key to creating a positive feedback loop between developers, publishers and end users. Open weights that become embedded in infrastructure — and the ecosystem position they create — are likely paths to commercial value.
Finally, with agents becoming the number one user type on the Hub, the next report’s readership and metrics may look very different: agents, not humans, may be driving growth and adoption.
Method notes
This analysis is based on activity observed on the Hugging Face Hub during the first seven months of 2026. The metrics used (downloads, likes, derivatives, model releases) reflect different aspects of ecosystem activity and should not be interpreted as direct measures of model quality, commercial adoption, or total market share. Downloads measure Hub usage but do not capture API activity, private deployments or distribution outside the Hub; likes capture attention; derivatives signal how much developers build on top of a model. Hub data is one perspective on open‑source AI adoption, not a complete measurement of the AI market.
