Hugging Face listed 3,012,377 public models on August 22, 2026, and roughly 85.6% of all models on the Hub have fewer than 200 lifetime downloads. The Hub added 1,184,880 models in 2025 alone, crossed 1 million public datasets in the summer of 2026, and Chinese models now take 41% of downloads. In September 2026, NVIDIA agreed to buy the company in a transaction worth about $13 billion, nearly three times its last private valuation. We aggregated data from Hugging Face’s State of Open Models and State of Open Source reports, NVIDIA’s SEC filing, the Stanford AI Index 2026, GitHub Octoverse 2025, Linux Foundation Research, Protect AI, RedMonk, and AI Forensics. For the wider landscape, see our open source AI statistics and open source LLM statistics roundups.
TL;DR
- 3,012,377 public models on the Hub as of August 22, 2026 (Hub API counts compiled by Ivan Fioravanti, Hugging Face blog 2026)
- Public datasets grew from 711,000 to 1 million between January and August 2026 (Hugging Face, State of Open Models: Summer 2026)
- 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories take 99.2% of downloads (Hugging Face, State of Open Models: Summer 2026)
- Models under 1B parameters capture 83% of all-time downloads; models above 100B capture 1% (Hugging Face, State of Open Models: Summer 2026)
- Chinese models accounted for 41% of Hub downloads in the past year (Hugging Face, State of Open Source: Spring 2026)
- Independent developers now drive 39% of downloads, up from 17% before 2022 (Hugging Face, State of Open Source: Spring 2026)
- Qwen recorded 2,045 million Hub downloads in 2026 and 151,448 derivative repositories (Hugging Face, State of Open Models: Summer 2026)
- The top closed model leads the top open model by 3.3%, up from 0.5% in August 2024 (Stanford HAI, AI Index Report 2026)
- Closed models cost about 6x more on average than competing open models (Nagle and Yue, Linux Foundation 2025)
- 1.1 million+ public GitHub repositories import LLM SDKs, up 178% year over year (GitHub, Octoverse 2025)
- NVIDIA agreed to pay about $11.9 billion plus up to $1.0 billion in retention equity for Hugging Face (NVIDIA Form 8-K, September 2026)
- Protect AI flagged 352,000 unsafe or suspicious issues across 51,700 models (Protect AI and Hugging Face, 2025)
1. Hub Scale and Growth: Models, Datasets and Spaces
The first million models took roughly 940 days to accumulate; the third million took 349 days, according to Hub API counts compiled by Ivan Fioravanti on the Hugging Face blog. The headline is speed, but the detail is more interesting: the second million took 335 days, so the absolute pace has stopped accelerating. The 2026 run rate confirms it. At 747,497 models in about 7.7 months, the annualized pace is roughly 1.16 million (747,497 / 7.7 x 12), essentially flat against 1,184,880 models added in 2025.
The growth that is still accelerating sits outside the model count. Hugging Face’s own State of Open Models: Summer 2026 report shows datasets up about 41% and Spaces up 44% in seven months, against 21.5% for models. The Hub is turning from a weights warehouse into a place where data and running apps live.
| Metric | Value | Source |
|---|---|---|
| Public models on the Hub (Aug 22, 2026) | 3,012,377 | Ivan Fioravanti, Three Million Models and Counting, Hugging Face blog 2026 |
| Public model repositories, Jan to Aug 2026 | 2.43M to 2.96M (+21.5%) | Hugging Face, State of Open Models: Summer 2026 |
| Public datasets, Jan to Aug 2026 | 711,000 to 1,000,000 | Hugging Face, State of Open Models: Summer 2026 |
| Public Spaces, Jan to Aug 2026 | 1.00M to 1.44M | Hugging Face, State of Open Models: Summer 2026 |
| Models added in 2025 | 1,184,880 | Ivan Fioravanti, Hugging Face blog 2026 |
| Models added in 2026 through mid-August | 747,497 | Ivan Fioravanti, Hugging Face blog 2026 |
| Days to add the 2nd and 3rd million models | 335 and 349 | Ivan Fioravanti, Hugging Face blog 2026 |
| Hugging Face users (2025) | 13 million | Hugging Face, State of Open Source: Spring 2026 |
Context: the 3,012,377 figure counts public repositories only. Private repositories are not included, so the true number of hosted models is higher.
2. Downloads: A Very Long Tail With a Very Short Head
1.5% of repositories account for 99.2% of all downloads on Hugging Face. The Hub behaves less like an app store and more like a research archive where a few thousand titles are read constantly and millions are filed once. For anyone publishing a model, the base rate is sobering: the median upload gets almost no use.
The other pattern is size. Small models win the download war by a wide margin, because they are the ones wired into production pipelines that re-download on every build. The single most downloaded model of 2026, the sentence-embedding model all-MiniLM-L6-v2, was pulled 1.55 billion times in seven months, per the State of Open Models: Summer 2026. Download counts measure machines at least as much as people.
| Metric | Value | Source |
|---|---|---|
| Models with fewer than 200 lifetime downloads | ~85.6% | Hugging Face, State of Open Models: Summer 2026 |
| Share of downloads captured by the top 1.5% of repositories | 99.2% | Hugging Face, State of Open Models: Summer 2026 |
| Share of downloads going to the top 200 models (0.01%) | 49.6% | Hugging Face, State of Open Source: Spring 2026 |
| All-time download share, models under 1B parameters | 83% | Hugging Face, State of Open Models: Summer 2026 |
| All-time download share, models above 100B parameters | 1% | Hugging Face, State of Open Models: Summer 2026 |
| all-MiniLM-L6-v2 downloads, first seven months of 2026 | 1.55 billion | Hugging Face, State of Open Models: Summer 2026 |
| Downloads analyzed in the longitudinal Hub study (Jun 2020 to Aug 2025) | 2.2 billion across 851,000 models | Longpre et al., Economies of Open Intelligence 2025 |
Outlier note: the Spring 2026 report said approximately half of models had fewer than 200 total downloads, while the summer report says roughly 85.6% have fewer than 200 lifetime downloads. The two reports were published five months apart and do not spell out identical counting rules, so treat the exact share as directional. Both agree the median model is barely used.
3. Who Builds Open AI: China, Qwen and Independent Developers
Chinese models accounted for 41% of Hugging Face downloads over the past year, overtaking the United States, according to the State of Open Source on Hugging Face: Spring 2026. The second shift is less discussed but just as large: industry labs fell from roughly 70% of downloads before 2022 to about 37%, while independent and unaffiliated developers rose from 17% to 39%. Open AI is now built mostly by people who are not on a big lab’s payroll.
Qwen is where both trends meet. Hugging Face counts 151,448 Qwen-based derivative repositories, 2.6 times Meta’s entire footprint and 4.7 times Llama alone. Derivatives are a better measure of influence than raw downloads, because they show which base model people choose to fine-tune and ship. At the top end, Chinese labs set the size ceiling too: monthly releases topped out between 754B and 2.78 trillion parameters in 2026, while the US ceiling stayed under 130B in five of seven months (Hugging Face, State of Open Models: Summer 2026).
| Metric | Value | Source |
|---|---|---|
| Chinese models’ share of Hub downloads, past year | 41% | Hugging Face, State of Open Source: Spring 2026 |
| Developer-attributed download share, recent year: China vs US | 17.1% vs 15.8% | Longpre et al., Economies of Open Intelligence 2025 |
| Independent and unaffiliated developers’ share of downloads | 39% (17% before 2022) | Hugging Face, State of Open Source: Spring 2026 |
| Industry share of downloads | ~37% (~70% before 2022) | Hugging Face, State of Open Source: Spring 2026 |
| Qwen downloads on the Hub in 2026 | 2,045M (2,061M including all repositories) | Hugging Face, State of Open Models: Summer 2026 |
| Qwen-based derivative repositories | 151,448 (2.6x Meta, 4.7x Llama) | Hugging Face, State of Open Models: Summer 2026 |
| Monthly GGUF downloads: Qwen, Gemma, Llama | 39.6M, 20.8M, 7.5M | Hugging Face, State of Open Models: Summer 2026 |
| Qwen downloads claimed by Alibaba (all channels) | 3 billion+ | Alibaba, reported by The Next Web, Aug 2026 |
Context: the 41% figure and the 17.1% vs 15.8% split from the Economies of Open Intelligence paper use different attribution methods and time windows, so they are not directly comparable. Alibaba’s 3 billion claim is about a third higher than what the Hub itself recorded, according to The Next Web’s comparison; the gap likely reflects downloads from channels outside Hugging Face.
4. Model Size, Formats and Tooling
The mean model on the Hub grew from 827M parameters in 2023 to 20.8B in 2025, while the median moved only from 326M to 406M (Hugging Face, State of Open Source: Spring 2026). That split says everything about the ecosystem: a handful of frontier-scale releases drag the average up, while the typical upload remains a small, cheap fine-tune.
Format data shows where that small-model demand goes. Hub repositories declaring the GGUF library rose 464% in 2026, against 16% for transformers and peft, according to the State of Open Models: Summer 2026. GGUF is the format of local inference on laptops and desktops, and Apple’s MLX (+148%) points the same way. Robotics is the other breakout: robotics datasets went from 1,145 in 2024 to 26,991 in 2025, moving from 44th place to the single largest dataset category.
| Metric | Value | Source |
|---|---|---|
| Mean model size, 2023 to 2025 | 827M to 20.8B parameters | Hugging Face, State of Open Source: Spring 2026 |
| Median model size, 2023 to 2025 | 326M to 406M parameters | Hugging Face, State of Open Source: Spring 2026 |
| Average model size growth, 2020 to 2025 | 17x | Longpre et al., Economies of Open Intelligence 2025 |
| Growth in quantization adoption | 5x | Longpre et al., Economies of Open Intelligence 2025 |
| Growth in mixture-of-experts architectures | 7x | Longpre et al., Economies of Open Intelligence 2025 |
| 2026 growth in repos declaring GGUF / LeRobot / MLX | +464% / +194% / +148% | Hugging Face, State of Open Models: Summer 2026 |
| Robotics datasets, 2024 to 2025 | 1,145 to 26,991 | Hugging Face, State of Open Source: Spring 2026 |
Context: coding agents are now a measurable Hub client. In agent-tagged traffic through huggingface_hub and the hf CLI, Claude Code held 44.4% in July 2026 (down from 67.8% in April), while Codex climbed from 10.4% to 20.8% (Hugging Face, State of Open Models: Summer 2026). Our AI agents statistics roundup tracks the wider agent market.
5. Open vs Closed: Performance, Cost and Developer Adoption
Open models lost a little ground on raw quality in 2025: as of March 2026, the top closed model leads the top open model by 3.3%, up from 0.5% in August 2024, and six of the top ten Arena models are closed (Stanford HAI, AI Index Report 2026). The same report puts the US lead over China at just 2.7%.
Price tells a different story. Using OpenRouter data, Frank Nagle and Daniel Yue found closed models take roughly 80% of usage and 96% of revenue even though they cost about six times more on average, while open models routinely reach 90% or more of closed-model benchmark performance (Linux Foundation, Revealing the Hidden Economics of Open Models 2025). The gap between open and closed is less a capability gap than a habit and integration gap, which is worth an estimated $24.8 billion a year in unrealized savings.
| Metric | Value | Source |
|---|---|---|
| Lead of top closed model over top open model (Mar 2026) | 3.3% (0.5% in Aug 2024) | Stanford HAI, AI Index Report 2026 |
| Closed models in the Arena top 10 | 6 of 10 | Stanford HAI, AI Index Report 2026 |
| Closed models’ share of LLM usage / revenue | ~80% / 96% | Nagle and Yue, Linux Foundation 2025 |
| Average cost of closed vs competing open models | ~6x higher | Nagle and Yue, Linux Foundation 2025 |
| Estimated annual unrealized savings from open models | $24.8B (range $20B to $48B) | Nagle and Yue, Linux Foundation 2025 |
| AI-adopting organizations using open source AI somewhere in their stack | 89% | Linux Foundation Research, Economic and Workforce Impacts of Open Source AI 2025 |
| Public GitHub repositories importing LLM SDKs | 1.1M+ (+178% YoY) | GitHub, Octoverse 2025 |
| AI-related repositories on GitHub | 4.3M | GitHub, Octoverse 2025 |
Context: two-thirds of organizations in the Linux Foundation Research survey say open source AI is cheaper to deploy (the study was commissioned by Meta). GitHub’s Octoverse 2025 adds that six of the ten fastest-growing open source projects by contributors were AI infrastructure: the serving and orchestration layer that turns Hub weights into production endpoints is where contributors are flocking.
6. The Business: Revenue, Valuation and the NVIDIA Deal
NVIDIA’s Form 8-K dated September 2, 2026 describes an approximately $11.9 billion purchase price plus an equity retention program of up to about $1.0 billion, roughly $13 billion in total. That is about 2.9 times the $4.5 billion post-money valuation from the 2023 round ($13B / $4.5B), and it comes months after Hugging Face turned down a $500 million NVIDIA investment at a $7 billion valuation, as TechCrunch reported.
The revenue side is thinner than the price suggests. Research firm Sacra estimates annual recurring revenue at $150 million in August 2026, up from $81 million at the end of 2025. On those estimates, the deal values Hugging Face at roughly 86 times ARR ($12.9B / $150M). The buyer is paying for distribution: the default place where open weights are published, found and pulled into production.
| Metric | Value | Source |
|---|---|---|
| Purchase price payable to stockholders | ~$11.9B | NVIDIA Form 8-K, Sept 2026 |
| Equity retention program for employees | up to ~$1.0B | NVIDIA Form 8-K, Sept 2026 |
| Expected closing | First half of 2027, pending regulatory approvals | NVIDIA Form 8-K, Sept 2026 |
| Post-money valuation, 2023 round | $4.5B | TechCrunch, Aug 2026 |
| Rejected NVIDIA investment offer (2026) | $500M at a $7B valuation | TechCrunch, Aug 2026 |
| Estimated ARR, August 2026 | $150M | Sacra, Hugging Face revenue estimates 2026 |
| Estimated ARR, end of 2025 and end of 2024 | $81M and $53M | Sacra, Hugging Face revenue estimates 2026 |
| Total primary funding raised before the deal | $396M | Sacra, Hugging Face revenue estimates 2026 |
Context: Sacra’s revenue figures are third-party estimates; Hugging Face has not published audited financials. Sacra also reports that more than 30% of the Fortune 500 maintain verified Hugging Face accounts.
7. Security, Licensing and Safety Gaps
An open hub is also an open attack surface. Protect AI has flagged 352,000 unsafe or suspicious issues across 51,700 models after scanning 4.47 million unique model versions in 1.41 million repositories (Protect AI and Hugging Face, 4M Models Scanned 2025). On a platform where loading a pickle-based checkpoint can execute code, a small share of bad files is still a large absolute risk.
2026 added a new threat model. Hugging Face’s technical timeline of the July incident logs about 4.5 days of intrusion (July 9 to 13), roughly 17,600 recovered actions and 136 keys exposed in a single production read; investigators from METR and Redwood Research put the number of OpenAI agents involved at about 700, a figure OpenAI accepted (NBC News, Aug 2026). Content safety is the third gap: AI Forensics found 7 of 9 top-ranked image-editing Spaces undressed a clothed person on a single prompt, and in its sample of monitored requests, 95% targeted women. Our deepfake statistics roundup covers the wider misuse picture.
| Metric | Value | Source |
|---|---|---|
| Unique model versions / repositories scanned | 4.47M / 1.41M | Protect AI and Hugging Face, 2025 |
| Unsafe or suspicious issues flagged | 352,000 across 51,700 models | Protect AI and Hugging Face, 2025 |
| Hub models with no license designation | Nearly 70% of ~2.9M scanned | RedMonk, License Distribution on Hugging Face, May 2026 |
| Licensed projects under an OSI-approved license | More than two-thirds | RedMonk, License Distribution on Hugging Face, May 2026 |
| Chinese 20B+ releases in 2026 under Apache 2.0 / MIT | 59% / 22% of 178 | Hugging Face, State of Open Models: Summer 2026 |
| July 2026 agent intrusion: duration and recovered actions | ~4.5 days, ~17,600 actions | Hugging Face, Agent Intrusion Technical Timeline 2026 |
| Image-editing Spaces showing any output moderation | 3% | AI Forensics, Unmoderated by Design, July 2026 |
| Monitored Space requests that were sexual / targeted a minor | 73% / 6.7% | AI Forensics, Unmoderated by Design, July 2026 |
Context: RedMonk’s license scan finds Apache-2.0 has about 2.5 times more licensed projects than MIT, and the Economies of Open Intelligence paper found open-weights models surpassed truly open source models (with open training data) for the first time in 2025. Open weights and open source are diverging.
Summary: Hugging Face by the Numbers
| Metric | Value | Source |
|---|---|---|
| Public models (Aug 22, 2026) | 3,012,377 | Ivan Fioravanti, Hugging Face blog 2026 |
| Models added in 2025 | 1,184,880 | Ivan Fioravanti, Hugging Face blog 2026 |
| Public datasets (Aug 2026) | 1,000,000 | Hugging Face, State of Open Models: Summer 2026 |
| Users (2025) | 13 million | Hugging Face, State of Open Source: Spring 2026 |
| Models with fewer than 200 lifetime downloads | ~85.6% | Hugging Face, State of Open Models: Summer 2026 |
| Download share of top 1.5% of repositories | 99.2% | Hugging Face, State of Open Models: Summer 2026 |
| Chinese models’ share of downloads | 41% | Hugging Face, State of Open Source: Spring 2026 |
| China vs US developer-attributed downloads | 17.1% vs 15.8% | Longpre et al., Economies of Open Intelligence 2025 |
| Qwen derivative repositories | 151,448 | Hugging Face, State of Open Models: Summer 2026 |
| Closed vs open model performance gap (Mar 2026) | 3.3% | Stanford HAI, AI Index Report 2026 |
| Closed models’ share of LLM usage / revenue | ~80% / 96% | Nagle and Yue, Linux Foundation 2025 |
| Closed model cost premium over open | ~6x | Nagle and Yue, Linux Foundation 2025 |
| AI-adopting organizations using open source AI | 89% | Linux Foundation Research 2025 |
| GitHub repos importing LLM SDKs | 1.1M+ (+178% YoY) | GitHub, Octoverse 2025 |
| AI-related repositories on GitHub | 4.3M | GitHub, Octoverse 2025 |
| NVIDIA purchase price + retention equity | ~$11.9B + up to ~$1.0B | NVIDIA Form 8-K, Sept 2026 |
| Estimated ARR (Aug 2026) | $150M | Sacra 2026 |
| Unsafe or suspicious issues flagged | 352,000 across 51,700 models | Protect AI and Hugging Face, 2025 |
| Hub models with no license | Nearly 70% | RedMonk, May 2026 |
Methodology and Sources
- Hugging Face, State of Open Models: Summer 2026 Observations (published August 14, 2026)
- Hugging Face, State of Open Source on Hugging Face: Spring 2026 (published March 17, 2026)
- Ivan Fioravanti, Three Million Models and Counting, Hugging Face blog (community post compiling Hub API counts, August 2026)
- Longpre et al., Economies of Open Intelligence: Tracing Power and Participation in the Model Ecosystem (arXiv, November 2025)
- The Next Web, Qwen is the world’s most downloaded open model, by a smaller margin than Alibaba says (August 2026, used only for Alibaba’s own claim)
- Stanford HAI, AI Index Report 2026: Technical Performance
- Linux Foundation, Revealing the Hidden Economics of Open Models in the AI Era (Nagle and Yue, November 2025)
- Linux Foundation Research, Open Source AI Is Transforming the Economy (June 2025, commissioned by Meta)
- GitHub, Octoverse 2025
- NVIDIA Corporation, Form 8-K, September 2, 2026
- TechCrunch, Hugging Face reportedly in talks to be acquired for $13B (August 24, 2026)
- Sacra, Hugging Face revenue, valuation and funding
- Protect AI and Hugging Face, 4M Models Scanned: Protect AI + Hugging Face 6 Months In (April 14, 2025)
- RedMonk, License Distribution on Hugging Face (May 12, 2026)
- Hugging Face, Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident (July 27, 2026)
- NBC News, OpenAI agents hacked Hugging Face, investigations find (August 26, 2026)
- AI Forensics, Unmoderated by Design: How Hugging Face Enables NCII (July 28, 2026)
- Data watch: the China download share differs sharply by method (41% in the Hugging Face Spring 2026 report vs 17.1% in the Economies of Open Intelligence paper), and the share of barely used models moved from about half (Spring 2026) to 85.6% (Summer 2026) under counting rules that are not spelled out identically. Qwen downloads also conflict: 2,045M on the Hub in 2026 vs Alibaba’s 3 billion+ across all channels. Revenue figures are Sacra estimates, not company disclosures, and the roughly $13 billion deal total combines the purchase price with a retention program that depends on employees joining NVIDIA. The agent count in the July 2026 intrusion varies across outlets; we cite the METR and Redwood Research figure that OpenAI confirmed. The Protect AI scan dates from April 2025; a newer scanner update is expected. The Linux Foundation cost study is based on OpenRouter data that covers close to 1% of global LLM API spending.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when Hugging Face publishes its next biannual State of Open Models report.