The era of large language models being locked inside industrial labs may have just cracked open. BLOOM, a 176-billion-parameter model trained on 46 natural languages and 13 programming languages, is now available for anyone to download, run, and study. The BigScience project behind it involved over 1,000 researchers from 70-plus countries and 250-plus institutions, making it the largest AI research collaboration ever assembled for a single model.
The training run itself took 117 days on the Jean Zay supercomputer near Paris, backed by a compute grant worth an estimated €3 million from French research agencies CNRS and GENCI. For languages like Spanish, French, and Arabic, BLOOM is the first model over 100 billion parameters ever built. That’s not a trivial milestone — it means researchers working in those languages finally have a foundation model at a scale previously reserved for English-dominant systems.
Access is unusually generous by LLM standards. Alongside the model weights, the team released intermediary checkpoints and optimizer states from training, something no major lab has done before. If you don’t have eight A100s lying around, an inference API backed by Google’s TPU cloud lets you prototype immediately through the Hugging Face Hub. The only gate is a Responsible AI License developed during the project itself, which governs how the model can be used and built upon.
The team is already pushing toward an instructable version, following the path of their earlier T0++ model, along with plans for more languages, a compressed variant with comparable performance, and use as a base for more complex architectures. That roadmap matters because open models have historically lagged their closed counterparts by months or years. BLOOM narrows that gap at a scale that was unthinkable even a year ago — and the fact that it will keep evolving as a living family of models rather than a one-off release is what could make it genuinely disruptive to the current dynamics of AI research.
