BLOOM (BigScience Large Open-science Open-access Multilingual Language Model) is an open-access, multilingual large language model developed by the BigScience project—a collaborative initiative involving over 1,000 researchers and coordinated by Hugging Face. BLOOM is designed to perform a wide range of natural language processing (NLP) tasks across 46 languages and 13 programming languages, making it one of the most inclusive and transparent LLMs available to date.

Built with a strong focus on openness, reproducibility, and multilingual capability, BLOOM is suitable for academic research, international development, public sector innovation, and responsible AI experimentation. It is freely available under the Responsible AI License (RAIL), with specific conditions aimed at preventing misuse.

What are the main features of BLOOM?

Multilingual capabilities across 46 human languages

BLOOM was trained to understand and generate text in a wide variety of natural languages, including low-resource and non-Western ones:

• Coverage includes major languages (English, French, Spanish, Arabic, Mandarin) and lesser-represented languages (Swahili, Vietnamese, Yoruba).

• Enables cross-lingual NLP applications, such as translation, summarization, and information retrieval.

• Supports inclusive AI research by enabling work on underrepresented linguistic communities.

This makes BLOOM highly relevant for global and multicultural use cases, especially in NGOs, academia, and language preservation efforts.

Support for 13 programming languages

In addition to natural language, BLOOM supports code generation and comprehension in multiple programming languages:

• Includes Python, JavaScript, C++, Rust, and more.

• Suitable for code completion, explanation, and teaching use cases.

• Can be used in multilingual environments that require code and documentation processing together.

This makes the model versatile not just for text, but also for technical workflows and educational contexts.

Open development with transparent training methodology

BLOOM was developed through a collaborative and transparent process:

• Trained on the ROOTS corpus, a publicly documented multilingual dataset with over 350 billion tokens.

• Built using public infrastructure (Jean Zay supercomputer in France) and open-source tools.

• Every decision, from dataset composition to model architecture, was publicly discussed and documented.

This approach enhances reproducibility, trust, and community involvement in large model development.

Ethical licensing through RAIL

Rather than permissive commercial licenses, BLOOM is distributed under a Responsible AI License (RAIL):

• Allows for broad access but restricts harmful uses (e.g., surveillance, disinformation).

• Encourages responsible research and innovation in AI.

• Aimed at balancing openness with ethical boundaries.

This license framework provides a middle ground between full openness and accountability.

Scalable architecture with multiple model sizes

BLOOM comes in several configurations, including the flagship 176B parameter model:

• Available in smaller versions (e.g., BLOOM-560M, BLOOM-3B) for resource-constrained users.

• Compatible with Hugging Face Transformers and major ML frameworks.

• Optimized for large-scale inference and research applications.

This makes BLOOM accessible to a wide range of users—from academic labs to enterprise teams—regardless of infrastructure limitations.

Why choose BLOOM?

• Multilingual by design: Built to support a diverse set of global languages, beyond English-centric models.

• Open and reproducible: Developed transparently with publicly available datasets and training logs.

• Code-aware and technically versatile: Supports both natural and programming languages for hybrid tasks.

• Responsible by license: Distributed under RAIL to ensure ethical and socially conscious usage.

• Scalable and accessible: Offers different model sizes to fit varied computational capacities.