What does open-weight mean?
Open-weight means the model weights are downloadable. It does not necessarily mean the training data, code or full development process is open, and the license may still restrict some uses.
Open-weight AI adoption, charted
Track the models, countries and local formats reshaping open AI. Every number is dated, sourced and scoped.
Sources: Vercel, Hugging Face and Stanford AI Index
Open weights now carry 57.6% of production tokens
Daily share on Vercel AI Gateway rose from 32.6% on 29 May to 57.6% on 26 August, peaking at 61.6%.
50% majority
Daily Vercel AI Gateway token volume Not global AI usage
View Vercel source| Date | Open weights | Closed weights |
|---|---|---|
| 2026-05-29 | 32.6% | 67.4% |
| 2026-05-30 | 33.7% | 66.3% |
| 2026-05-31 | 34.4% | 65.6% |
| 2026-06-01 | 26.8% | 73.2% |
| 2026-06-02 | 28.8% | 71.2% |
| 2026-06-03 | 29.9% | 70.1% |
| 2026-06-04 | 32.1% | 67.9% |
| 2026-06-05 | 32.5% | 67.5% |
| 2026-06-06 | 43.9% | 56.1% |
| 2026-06-07 | 42.3% | 57.7% |
| 2026-06-08 | 35.1% | 64.9% |
| 2026-06-09 | 36.3% | 63.7% |
| 2026-06-10 | 33.2% | 66.8% |
| 2026-06-11 | 34.3% | 65.7% |
| 2026-06-12 | 36.2% | 63.8% |
| 2026-06-13 | 44.5% | 55.5% |
| 2026-06-14 | 41.5% | 58.5% |
| 2026-06-15 | 33.6% | 66.4% |
| 2026-06-16 | 38.2% | 61.8% |
| 2026-06-17 | 38.4% | 61.6% |
| 2026-06-18 | 29.8% | 70.2% |
| 2026-06-19 | 26.8% | 73.2% |
| 2026-06-20 | 30.0% | 70.0% |
| 2026-06-21 | 29.9% | 70.1% |
| 2026-06-22 | 30.8% | 69.2% |
| 2026-06-23 | 30.8% | 69.2% |
| 2026-06-24 | 28.4% | 71.6% |
| 2026-06-25 | 30.2% | 69.8% |
| 2026-06-26 | 33.0% | 67.0% |
| 2026-06-27 | 38.3% | 61.7% |
| 2026-06-28 | 35.6% | 64.4% |
| 2026-06-29 | 34.9% | 65.1% |
| 2026-06-30 | 34.1% | 65.9% |
| 2026-07-01 | 27.5% | 72.5% |
| 2026-07-02 | 31.3% | 68.7% |
| 2026-07-03 | 41.4% | 58.6% |
| 2026-07-04 | 43.0% | 57.0% |
| 2026-07-05 | 46.6% | 53.4% |
| 2026-07-06 | 44.6% | 55.4% |
| 2026-07-07 | 46.7% | 53.3% |
| 2026-07-08 | 44.2% | 55.8% |
| 2026-07-09 | 46.3% | 53.7% |
| 2026-07-10 | 46.5% | 53.5% |
| 2026-07-11 | 50.7% | 49.3% |
| 2026-07-12 | 49.8% | 50.2% |
| 2026-07-13 | 46.4% | 53.6% |
| 2026-07-14 | 49.0% | 51.0% |
| 2026-07-15 | 45.7% | 54.3% |
| 2026-07-16 | 44.5% | 55.5% |
| 2026-07-17 | 42.6% | 57.4% |
| 2026-07-18 | 46.3% | 53.7% |
| 2026-07-19 | 49.6% | 50.4% |
| 2026-07-20 | 49.1% | 50.9% |
| 2026-07-21 | 51.0% | 49.0% |
| 2026-07-22 | 40.6% | 59.4% |
| 2026-07-23 | 42.2% | 57.8% |
| 2026-07-24 | 49.0% | 51.0% |
| 2026-07-25 | 49.3% | 50.7% |
| 2026-07-26 | 54.1% | 45.9% |
| 2026-07-27 | 48.3% | 51.7% |
| 2026-07-28 | 47.6% | 52.4% |
| 2026-07-29 | 45.4% | 54.6% |
| 2026-07-30 | 50.0% | 50.0% |
| 2026-07-31 | 45.8% | 54.2% |
| 2026-08-01 | 52.8% | 47.2% |
| 2026-08-02 | 53.5% | 46.5% |
| 2026-08-03 | 49.0% | 51.0% |
| 2026-08-04 | 48.0% | 52.0% |
| 2026-08-05 | 48.6% | 51.4% |
| 2026-08-06 | 54.1% | 45.9% |
| 2026-08-07 | 56.7% | 43.3% |
| 2026-08-08 | 61.4% | 38.6% |
| 2026-08-09 | 60.2% | 39.8% |
| 2026-08-10 | 54.0% | 46.0% |
| 2026-08-11 | 53.6% | 46.4% |
| 2026-08-12 | 50.5% | 49.5% |
| 2026-08-13 | 52.7% | 47.3% |
| 2026-08-14 | 52.3% | 47.7% |
| 2026-08-15 | 50.8% | 49.2% |
| 2026-08-16 | 55.8% | 44.2% |
| 2026-08-17 | 52.5% | 47.5% |
| 2026-08-18 | 52.7% | 47.3% |
| 2026-08-19 | 57.0% | 43.0% |
| 2026-08-20 | 57.3% | 42.7% |
| 2026-08-21 | 57.0% | 43.0% |
| 2026-08-22 | 61.6% | 38.4% |
| 2026-08-23 | 60.7% | 39.3% |
| 2026-08-24 | 50.1% | 49.9% |
| 2026-08-25 | 52.5% | 47.5% |
| 2026-08-26 | 57.6% | 42.4% |
The geography of open AI
Where models are built and where they are adopted do not always line up. Production and usage need separate measures.
2. The U.S. still produces more notable models
In 2025, U.S.-based institutions produced 59 notable models, compared with 35 from China.
Notable models released in 2025 Stanford AI Index 2026
View source| Country | Models |
|---|---|
| United States | 59 |
| China | 35 |
3. China leads the open download layer
Chinese-origin models represented 41% of Hugging Face downloads over the past year.
41%
59%
Hugging Face downloads Unclear country origins are a known limitation
View source| Origin | Share |
|---|---|
| China | 41% |
| Rest of world | 59% |
Production leadership and adoption leadership are not the same metric.
The practical layer
4. Small models dominate real downloads
Among Hugging Face models that declare a parameter count, models under 1B represent 83% of all-time downloads.
83%Under 1B
Hugging Face downloads Declared parameter counts only
View source| Model size | Share |
|---|---|
| Under 1B | 83% |
| 1B to 100B | 16% |
| Above 100B | 1% |
5. Local inference formats are growing faster
During the first seven months of 2026, repositories declaring GGUF grew far faster than the Hub overall.
Repository growth, Jan–Jul 2026 Hugging Face
View source| Format or group | Growth |
|---|---|
| GGUF | 464% |
| MLX | 148% |
| Hub overall | 21.5% |
| Transformers / PEFT | 16% |
Read the numbers correctly
Vercel token shares reflect production traffic routed through AI Gateway, not all global AI use. Downloads show activity inside one ecosystem. They are not unique users, revenue or global market share. Country refers to the originating organization, not where inference runs.
Open-weight AI statistics, explained
Is 57.6% global AI usage?
No. It is the daily share of tokens routed through Vercel AI Gateway on models with downloadable weights. It is a production signal inside that gateway, not global market share.
Why focus on small models and local formats?
Downloads and repository growth show where models become practical. GGUF and MLX connect open weights to hardware people can actually operate.