AI · Open-Source AI
HuggingFace Models
Public ML models hosted on HuggingFace: the “GitHub for AI” curve no closed lab can capture.
- ~4.8K (2020) → 3M models (2026)
- Crossed 1M models on 26 Sep 2024
- At least 1 in 6 tagged fine-tune or quantised
- Pre-2020 figures are a different population
HuggingFace grew out of the Transformers library it released in 2018, and opened its public model hub, where anyone can upload and share weights, around 2020. It crossed 1 million public models on 26 September 2024 and hosts just over 3.1 million as of October 2026 — roughly 660× the first archived hub-wide count of 4,822 at the end of 2020. The tiny 2018–19 figures are the curated launch checkpoints of the Transformers library, a different population, so a multiplier from there would compare two different things.
Where the training-compute and inference-price curves here follow the closed frontier, this one counts the open-source explosion: fine-tunes, LoRAs, quantisations, multimodal variants, and whatever other models people upload. A 600B-parameter language model and a 30 MB embedding model sit on the same shelf. Hub metadata tags at least one in six models as a fine-tune or quantisation of another. The true derivative share is higher but unknowable, since the relationship metadata is optional.
| Year | Value | Note | Projected |
|---|---|---|---|
| 2018 | 5 | Transformers library released | |
| 2019 | 71 | ||
| 2020 | 4,822 | ||
| 2021 | 24,645 | ||
| 2022 | 108,473 | ||
| 2023 | 447,557 | ||
| September 2024 | 1,000,000 | 1M models (26 Sep 2024) | |
| 2025 | 2,366,714 | ||
| October 2026 | 3,172,087 | ~3.17M models, 20 Oct 2026 |