A 176-billion parameter open multilingual language model, built collaboratively by over 1,000 researchers. The first proof at scale that large language models don't require corporate labs — just public compute and collective will.
What It Is
BLOOM (BigScience Large Open-science Open-access Multilingual Language Model) was trained through BigScience, a year-long research workshop coordinated by Hugging Face. Unlike corporate models:
- Truly open — weights, code, and training data disclosed
- Multilingual — 46 natural languages plus 13 programming languages
- Collaborative — 1,000+ researchers from 60+ countries
- Documented — model cards, data documentation, training details all public
BLOOM was one of the first demonstrations that large language models could be built outside corporate labs.
Why It Matters for Redistribution
Who benefits:
- Researchers — can study, modify, and build on a frontier-scale model
- Non-English speakers — 46 languages vs. English-dominant commercial models
- Global South — researchers could participate in training, not just consume results
What changed: BLOOM proved open, collaborative AI development is possible at scale. It established precedents for:
- International research coordination
- Transparent training data curation (ROOTS dataset)
- Responsible AI licensing (RAIL license)
- Public compute allocation (French supercomputer Jean Zay)
Who funded it: French government (GENCI/IDRIS) provided compute via Jean Zay supercomputer. Hugging Face coordinated. No single corporate interest controlled development.
The bigger picture: BLOOM stands as a counterexample to The Frontier Lab Oligopoly. When Compute Is the New Oil, a publicly funded supercomputer training an open model is the equivalent of a public water utility — infrastructure for the commons rather than private extraction. The ROOTS dataset, with its documented provenance, also addresses the questions raised in Who Owns the Training Data by making curation decisions transparent and accountable.
The RAIL (Responsible AI License) represents an experiment in Open Weights Are Not Open Source territory — more open than proprietary, more governed than permissive. It's an attempt to balance access with accountability.
The tensions:
- Model is now dated relative to frontier
- Maintenance and continued development remain challenges
- RAIL license is novel but not battle-tested
- Coordination costs were high
Current Status
Active — BLOOM weights remain available and widely used. Development has slowed, but the model serves as foundation for fine-tuned variants and research.
The BigScience approach influenced subsequent open efforts including OpenLLaMA, Falcon, and continued Hugging Face initiatives.
How to Get Involved
- Use it: Access weights through Hugging Face
- Research: Study the training process and documentation
- Build: Create fine-tuned variants for specific languages or domains
- Learn: The BigScience documentation is a masterclass in open ML
Related
- Hugging Face
- EleutherAI
- Open Source vs. Open Weights
- Open Weights Are Not Open Source
- Compute Is the New Oil
- Who Owns the Training Data
- Datasheets for Datasets
- Public AI
Sources
- BigScience Workshop. "BLOOM." bigscience.huggingface.co
- Scao, Teven Le et al. "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model." arXiv, 2022. arXiv
- Laurençon, Hugo et al. "The BigScience ROOTS Corpus." NeurIPS Datasets Track, 2022. arXiv