Introducing The World's Largest Open Multilingual Language Model: BLOOM
Hugging Face ● Covered by 2 sources
Hugging Face and 1000+ researchers just released BLOOM, a 176-billion-parameter language model that's fully open. Unlike GPT-3 style models locked behind corporate doors, anyone can download, inspect, and run this one.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For years, the biggest language models have lived behind closed doors, guarded by the handful of companies with the GPUs and cash to train them. BLOOM changes that math. Built by a coalition of over 1000 researchers spanning more than 70 countries and 250-plus institutions, this 176-billion-parameter model is now sitting on Hugging Face's servers, free for anyone willing to accept its Responsible AI License.
The scale of the collaboration is almost as notable as the model itself. Training ran for 117 days, from March 11 to July 6, on the Jean Zay supercomputer outside Paris, powered by a roughly €3 million compute grant from French agencies CNRS and GENCI. That's public infrastructure doing what usually only Big Tech budgets can afford.
What makes BLOOM genuinely different isn't just its size — it's the breadth. The model speaks 46 natural languages and 13 programming languages, and for most of those, including Spanish, French, and Arabic, this marks the first time a language model north of 100 billion parameters has existed at all. English-centric AI has dominated this space for so long that a proper multilingual heavyweight feels overdue.
Hugging Face isn't just handing over the finished weights, either. They're releasing intermediary checkpoints and optimizer states from training, letting researchers pick apart how the model evolved rather than just poke at the final product. And for anyone lacking a rack of A100 GPUs, an inference API running on Google's TPU cloud lets curious users test-drive BLOOM without any hardware investment.
The team behind it, the BigScience workshop, frames this as a starting point rather than a finished product. Plans include instruction-tuning similar to their earlier T0++ effort, adding more languages, and shrinking the model without sacrificing performance. BLOOM, in their words, is meant to be a living family of models — one the wider community gets to help raise.
My take — AI-written commentary, not fact-checked reporting
This is the release that actually matters more than another flashy chatbot demo — a transparent, multilingual, community-built alternative to the walled gardens at OpenAI and Google. I'm skeptical BLOOM will match GPT-3 on raw benchmarks, but that's not really the point; open weights and public checkpoints let smaller labs actually study how these systems work instead of guessing from an API. Europe funding this through public compute grants is the kind of move I wish happened more often instead of just regulating from the sidelines.
Read more about this at: Hugging Face