83% of Model Downloads Are Under 1B
The model everyone writes about and the model everyone actually runs are not the same model.
Founder, Automation Squad ·
The short answer
Hugging Face's State of Open Models report, published August 14, 2026, finds that models under 1B parameters account for 83% of all-time downloads on the Hub, while models above 100B account for 1%. It also finds extreme concentration: 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.
The numbers, and what they mean for what you actually deploy
Every figure below is from Hugging Face's own report. The interesting ones are not the growth numbers, they are the distribution numbers.
Look at what you are actually calling in production
List your live workloads and the model each one uses. Most teams find at least one job running on a large model because that is what was current when it was built, not because the job needs it.
Pick the single job with the highest volume and the lowest complexity
Classification, extraction, tagging, routing, short summarisation. These are where the download data is pointing — high frequency, low reasoning requirement.
Test it against a small model, and measure rather than eyeball
Run the same fifty real inputs through both. Count exact disagreements, not vibes. If the small model matches on forty-eight, the question becomes whether the two edge cases are worth the price difference.
Price the difference over a month, not a call
Per-request cost differences look trivial and monthly cost differences do not. Multiply by your real volume before deciding.
Treat the concentration figure as a warning about picking
With 85.6% of models under 200 lifetime downloads, most of what is on the Hub is untested by anyone. Popularity is not quality, but on an open hub it is the cheapest available proxy for 'someone else has hit the obvious bugs first'.
| Measure | Figure |
|---|---|
| Downloads — models under 1B params | 83% of all-time |
| Downloads — models above 100B params | 1% of all-time |
| 2026 downloads — models above 70B | 3% of annual volume |
| Public model repos | 2.43M → 2.96M |
| Datasets | 711K → 1M |
| Spaces | 1.00M → 1.44M |
| Qwen-derivative repos | 151,448 |
| Google-derivative repos | 82,506 |
| Meta / Llama total footprint | ~58,000 |
| Models with under 200 lifetime downloads | 85.6% |
| Share of downloads from the top 1.5% of repos | 99.2% |
Take it with you
STATE OF OPEN MODELS — SUMMER 2026 (Hugging Face) Published Aug 14, 2026 Source: https://automationsquad.com/news/hugging-face-open-models-summer-2026/ DOWNLOAD SHARE BY SIZE Models under 1B params ........ 83% of all-time downloads Models above 100B params ...... 1% of all-time downloads 2026 downloads above 70B ...... 3% of annual volume HUB GROWTH (Jan -> Aug 2026) Public model repos ............ 2.43M -> 2.96M Datasets ...................... 711K -> 1M Spaces ........................ 1.00M -> 1.44M DERIVATIVE ECOSYSTEMS Qwen .......................... 151,448 repos Google ........................ 82,506 repos Meta / Llama .................. ~58,000 repos CONCENTRATION Models with <200 lifetime downloads ....... 85.6% Share of downloads from top 1.5% of repos . 99.2% WHAT TO DO WITH IT [ ] List live workloads + the model each one uses [ ] Pick the highest-volume, lowest-complexity job [ ] Run 50 real inputs through it AND a small model [ ] Count exact disagreements — don't eyeball it [ ] Price the difference over a MONTH at real volume [ ] Remember: 85.6% of models are effectively untested by anyone
Hugging Face published its State of Open Models report for Summer 2026 on August 14, covering the first seven months of the year on the Hub.
The facts: 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. On download share by size, models under 1 billion parameters account for 83% of all-time downloads while models above 100 billion account for 1%; restricted to 2026 alone, models above 70 billion parameters make up 3% of the year's volume. Qwen derivatives now number 151,448 repositories against 82,506 for Google and roughly 58,000 for Meta's total Llama footprint. And the distribution is severely concentrated: 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads.
Automation Squad's take: there is a permanent gap between the model that gets written about and the model that gets deployed, and this is the clearest measurement of it anyone publishes. Frontier launches dominate coverage because they are genuinely interesting; sub-1B models dominate downloads because they are what fits on the hardware, inside the latency budget, and within the invoice. Neither group is wrong. But if you have been choosing models by reading launch coverage, the data says you are probably reaching for something larger than your job needs. The concentration figure deserves equal attention for the opposite reason — most of what is on the Hub has been tried by almost nobody, so an obscure model is not an undiscovered gem by default.
Run this now: list your live AI workloads and write the model each one uses next to it. Find the one with the highest volume and the lowest complexity — classification, extraction, routing, tagging — and run fifty real inputs through both your current model and a small one. Count the exact disagreements rather than forming an impression. Then price the difference across a month at your real volume, because per-call differences look negligible and monthly ones rarely do.
By the numbers
The data behind the story — figures from cited sources, with our own analysis labelled.
All-time downloads on the Hugging Face Hub. The gap between what gets covered and what gets pulled is the story.
Questions people are asking
- What size model do people actually download most?
- Small ones. Hugging Face's Summer 2026 report puts models under 1B parameters at 83% of all-time downloads, against 1% for models above 100B. Looking at 2026 alone, models above 70B account for 3% of the year's download volume.
- How big is the Hugging Face Hub now?
- 2.96 million public model repositories, up from 2.43 million at the start of 2026. Datasets grew from 711,000 to 1 million and Spaces from 1.00 million to 1.44 million over the same period.
- Which base model has the largest derivative ecosystem?
- Qwen, at 151,448 repositories — ahead of Google's 82,506 and Meta's total Llama footprint of roughly 58,000.
- Does a high download count mean a model is good?
- No, but on a hub where 85.6% of models have under 200 lifetime downloads, it is a reasonable proxy for whether anyone has actually stress-tested it. The concentration is severe: 1.5% of repositories account for 99.2% of all downloads.
Further reading
- Hugging Face — State of Open Models: Summer 2026 — The full report, including the breakdowns we did not pull out here.
Last checked August 14, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].