AIWiki
Malaysia
Back to all articles
AI FoundationsBLOOMmultilingual LLMopen science

BLOOM

3 min readUpdated November 2026
BLOOM
Developer
BigScience research workshop (coordinated by Hugging Face)
Released
12 July 2022
Parameters
176 billion
Languages
46 natural, 13 programming
Licence
Apache 2.0 (code); BigScience RAIL v1.0 (weights)
Related
GPT-3, Llama, OLMo, open-weight models
BLOOM (BigScience Large Open-science Open-access Multilingual Language Model) is a 176-billion-parameter transformer language model released in July 2022 as the flagship output of BigScience, a volunteer-driven research workshop created to provide a transparently documented alternative to proprietary AI models.[1] Unlike GPT-3 and its successors, BLOOM was trained from the outset to generate text in 46 natural languages and 13 programming languages.

History and Development

BigScience ran as a one-year research workshop coordinated by Hugging Face, with funding from the French government and several hundred volunteer researchers and engineers from academia and industry.[2] The model was trained between March and July 2022 on the Jean Zay public supercomputer in France, operated by GENCI and IDRIS (CNRS), using the open-source Megatron-DeepSpeed stack.[3]

BLOOM's training corpus, ROOTS, combines the OSCAR web scrape (about 38% of the total) with data drawn from a manually selected, documented list of language sources — an explicit attempt to make the provenance of training data auditable. In total the model was trained on roughly 366 billion tokens (about 1.6 TB of text).[1] The source code was released under Apache 2.0, while the model weights carry BigScience's Responsible AI License (RAIL) v1.0, which grants open access and reuse rights subject to usage restrictions on harmful applications.[4]

Significance

BLOOM was the first open model of GPT-3 scale to make multilingual generation a first-class design goal rather than an incidental capability, and it demonstrated that a large model could be built and documented collaboratively, with dataset sheets, evaluation reports and training logs published alongside the weights.[2] It became the basis for downstream products including BLOOMChat and the original HuggingChat assistant, and its open-by-default methodology directly informed later open-model programmes such as EleutherAI's GPT-NeoX, Allen AI's OLMo and the general open-weights movement.[3]

Because BLOOM is openly downloadable, researchers can study the behaviour of a frontier-scale model — its biases, failure modes and language coverage — without an API, which is central to AI evaluation, auditability and red-teaming work.[4]

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: BLOOM matters to Malaysia chiefly as a template for multilingual AI built outside the Anglophone labs. Bahasa Melayu and the country's Chinese- and Tamil-language communities are poorly served by English-first proprietary models, and BLOOM showed that open, collaboratively built corpora are the practical route to closing that gap — the approach taken by regional and local efforts such as SEA-LION and Malaysia's own ILMU and MALLaM language models.[5]

Malaysian universities and research groups (UM, USM, UTAR, AI universities' NLP labs) can download and inspect BLOOM free of charge, enabling coursework, fine-tuning experiments and bias studies that would otherwise require paid API credits — significant in a market where MDEC and the National AI Office are pushing AI literacy and where HRD Corp-funded training needs non-proprietary teaching material. Its RAIL licence also offers a model for Malaysian organisations wanting open weights with usage conditions, relevant to PDPA and responsible-AI expectations. At national level, BLOOM's documented-data argument reinforces Malaysia's position that sovereign AI capability requires curated, well-governed multilingual datasets rather than dependence on foreign model APIs.

References

  1. ↑Wikipedia. (2022). BLOOM (language model). https://en.wikipedia.org/wiki/BLOOM_(language_model)
  2. ↑Hugging Face. (2022). BLOOM: a 176B-parameter open multilingual language model (project announcement). https://huggingface.co/blog/bloom
  3. ↑Hugging Face. (2022). BLOOM: a 176B-parameter open multilingual language model. https://huggingface.co/bigscience/bloom
  4. ↑BigScience. Responsible AI License (RAIL) v1.0. https://huggingface.co/bigscience/bloom/blob/main/README.md
  5. ↑Malaysia Digital Economy Corporation. AI and language technology programmes. https://www.mdec.my/
Back to all articles