이미지: METAL LAB 생성
Summary
- Ai2 has expanded its partnership with Hugging Face, increasing its storage capacity on the Hub roughly threefold to about 2 petabytes
- The lab is also no longer subject to the Hub's standard download speed limits, allowing large datasets and multi-checkpoint models to be downloaded at maximum speed
- Ai2's models and datasets are reported to have been downloaded more than 50 million times on the Hugging Face Hub since spring 2024
- 발표일
- 2026년 8월 6일
- 저장용량 확대
- 기존 대비 약 3배, 약 2페타바이트
- 다운로드 제한
- 허브 표준 rate limit 적용 제외
- 누적 다운로드
- 2024년 봄 이후 5000만 회 이상
- 공개 규모
- 허깅페이스 기준 모델 900개 이상, 데이터셋 1200개 이상 (Clem Delangue 언급)
- 허브 내 게시 순위
- 허깅페이스 공식 히트맵 기준 연간 신규 아티팩트 게시량 최다 기관
- 대표 모델·벤치마크
- Olmo, Tülu 3, Molmo, OlmoEarth, AstaBench, olmOCR-Bench
Why storage capacity tripled
The Allen Institute for AI (Ai2) has expanded its collaboration with Hugging Face, roughly tripling its storage capacity on the Hub to about 2 petabytes. Two petabytes is equivalent to thousands of standard consumer hard drives combined — enough space to upload entire training checkpoints and raw datasets for large language models with room to spare. On top of that, Ai2 is no longer subject to the Hub's standard download speed limits. Previously, large datasets or models composed of multiple checkpoints took time to download due to speed restrictions, but now they can be downloaded at maximum speed.
Behind this move lies demand that can be measured in hard numbers. Ai2's models and datasets are reported to have been downloaded more than 50 million times on the Hugging Face Hub since spring 2024. According to Hugging Face's own heatmap data, Ai2 also ranks as the institution with the highest volume of new artifacts posted annually among those the platform tracks.
What is Ai2
Ai2 is a nonprofit AI research lab founded by Microsoft co-founder Paul Allen. Unlike typical "open-source" projects that release only model weights, Ai2 has championed a "fully open" approach — publishing training scripts, intermediate checkpoints, data mixture ratios, evaluation results, and even failed experiments (ablations). This approach allows other researchers to reproduce and verify exactly how a model was built, setting it apart from open models that release weights alone.
What Ai2 has released
The materials Ai2 has uploaded to Hugging Face span its Olmo language model family, the post-training model Tülu 3, the multimodal model Molmo, the Earth-observation model OlmoEarth, and AstaBench, an evaluation tool for scientific research agents.
| Project | Description |
|---|---|
| Olmo | Fully open language model |
| Tülu 3 | Post-training model |
| Molmo | Multimodal model |
| OlmoEarth | Earth-observation AI |
| AstaBench | Benchmark for evaluating scientific research agents |
| olmOCR-Bench | Document recognition (OCR) benchmark |
Among these, olmOCR-Bench is a benchmark that evaluates how well models can read multi-column layouts, tables, handwriting, and captioned charts within documents. The Hugging Face team has adopted it as their default OCR benchmark and integrated it with the Hub's leaderboard feature, allowing anyone to compare document-understanding model performance in one place.
Ai2 interim CEO Peter Clark said, "Releasing everything that goes into building a model, not just the model itself, is core to how Ai2 works." This statement suggests that the expanded partnership goes beyond simple infrastructure support — it's a practical measure to uphold the lab's open-research approach.
What Hugging Face gains
Through this expansion, Hugging Face is emphasizing that its platform has become infrastructure supporting open research across diverse fields — from Ai2's vision-language-action model for robotics, MolmoAct, to its climate-focused OlmoEarth model. Hugging Face functions as a community hub for sharing open-source models and datasets, having established itself as the de facto standard platform for AI researchers and developers to download and deploy models. Since the more artifacts major labs contribute, the greater the Hub's presence grows, this partnership can be seen as a decision that aligns the interests of both sides.
So what changes
For researchers and developers, the practical impact is clear. Previously, downloading large checkpoints or massive training datasets took time due to speed restrictions, but now they can access all of Ai2's published materials without that bottleneck. This also means more materials are becoming available that let people examine the entire training process, rather than open projects that hand over only a single set of model weights. As seen recently in NVIDIA's open world model Cosmos 3 and in comments about Cohere's open-source strategy, major AI companies are competing while each defining "openness" differently — and Ai2's latest expansion can be viewed as a case of reinforcing the standard for openness at the infrastructure level.
