Open weights, counted: the data
Whose language models does the Hugging Face Hub actually serve? The count below credits each model’s downloads to the lab whose weights it was built from, so a quantized or fine-tuned copy of a Qwen or Llama model counts for Qwen or Llama. 2024-07 to 2026-09.
1. Share of language-model downloads, by lab
Qwen: 7.4% in 2024-07, 50.6% in 2026-09. Meta peaked at 29.8% in 2024-11 and was at 4.6% in 2026-09.
2. The market grew as the shares moved
Monthly mean of the 30-day count, 2024-07 101M, 2025-09 315M, 2026-09 883M. Meta’s own downloads fell by about a fifth over the last year while the market multiplied by 2.8. Jumps (autumn 2024, spring 2026) are in the source data and the desk has not explained them.
3. The largest labs, 2026-09
4. How much of a lab’s count is other people’s copies
The share of each lab’s attributed downloads that land on derivatives (quantizations, fine-tunes, adapters, merges) published by other accounts. A high share means the lab’s weights are circulating mostly through others.
5. The top 40 labs
| # | Lab | 2026-09 (M) | Share | 2025-09 (M) | Change |
|---|---|---|---|---|---|
| 1 | Alibaba (Qwen) | 446.3 | 50.6% | 92.6 | x4.82 |
| 2 | 93.6 | 10.6% | 25.6 | x3.66 | |
| 3 | Meta | 40.8 | 4.6% | 50.7 | x0.81 |
| 4 | OpenAI | 31.1 | 3.5% | 28.9 | x1.08 |
| 5 | DeepSeek | 29.7 | 3.4% | 16.1 | x1.85 |
| 6 | ornith-ai | 27.9 | 3.2% | 0.0 | new |
| 7 | NVIDIA | 15.5 | 1.8% | 1.9 | x8.33 |
| 8 | Zhipu (Z.ai) | 14.4 | 1.6% | 1.6 | x8.77 |
| 9 | Microsoft | 10.6 | 1.2% | 9.8 | x1.07 |
| 10 | Hugging Face | 10.0 | 1.1% | 3.3 | x3.03 |
| 11 | EleutherAI | 8.4 | 0.9% | 1.5 | x5.65 |
| 12 | Mistral | 7.0 | 0.8% | 6.7 | x1.05 |
| 13 | Moonshot | 6.4 | 0.7% | 0.9 | x6.82 |
| 14 | farbodtavakkoli | 6.4 | 0.7% | 0.0 | new |
| 15 | openbmb | 5.8 | 0.7% | 0.6 | x9.00 |
| 16 | 01-ai | 5.7 | 0.6% | 5.4 | x1.05 |
| 17 | LiquidAI | 5.4 | 0.6% | 0.3 | x15.90 |
| 18 | datalab-to | 5.3 | 0.6% | 0.0 | new |
| 19 | baidu | 4.2 | 0.5% | 1.8 | x2.29 |
| 20 | ibm-granite | 3.8 | 0.4% | 1.0 | x3.96 |
| 21 | RadixArk | 3.6 | 0.4% | 0.0 | new |
| 22 | MiniMaxAI | 3.3 | 0.4% | 0.2 | x20.44 |
| 23 | llava-hf | 3.3 | 0.4% | 2.3 | x1.42 |
| 24 | thinkingmachines | 3.3 | 0.4% | 0.0 | new |
| 25 | OpenGVLab | 3.1 | 0.3% | 4.8 | x0.64 |
| 26 | allenai | 2.7 | 0.3% | 0.5 | x5.30 |
| 27 | HauhauCS | 2.2 | 0.2% | 0.0 | new |
| 28 | distilbert | 2.1 | 0.2% | 3.8 | x0.55 |
| 29 | meta-models | 1.9 | 0.2% | 0.0 | new |
| 30 | deepreinforce-ai | 1.9 | 0.2% | 0.0 | new |
| 31 | TinyLlama | 1.9 | 0.2% | 2.4 | x0.78 |
| 32 | vikhyatk | 1.8 | 0.2% | 0.2 | x8.41 |
| 33 | Jackrong | 1.8 | 0.2% | 0.0 | new |
| 34 | DavidAU | 1.6 | 0.2% | 0.2 | x7.86 |
| 35 | stepfun-ai | 1.5 | 0.2% | 0.1 | x9.93 |
| 36 | bigscience | 1.4 | 0.2% | 2.2 | x0.67 |
| 37 | LGAI-EXAONE | 1.4 | 0.2% | 1.0 | x1.40 |
| 38 | apple | 1.4 | 0.2% | 0.5 | x2.54 |
| 39 | tencent | 1.4 | 0.2% | 0.1 | new |
| 40 | ByteDance-Seed | 1.3 | 0.1% | 2.3 | x0.55 |
Method and limits
Each model in the Model Pulse table is followed up its base-model links to a root, and its downloads are credited to the root’s author, grouped into labs (for example Meta = meta-llama and facebook; Qwen = Qwen and Alibaba’s other accounts). Lineage is as of the latest snapshot, applied to the whole history. Models with no stated base count for their own author. Downloads of models the source does not track (under 10 downloads in 30 days, 50 all-time, and no likes) are not counted. Some days are missing in the source (August 2024, June 2025, April 2026, May and June 2026); monthly means use the days present. The desk did not verify the source against the Hub.
See the editorial: Open Weights, Counted.