BLOOMZ

BLOOMZ adapts the BLOOM family into a general instruction follower by fine-tuning on cross-task, cross-lingual supervision so the model learns to read a prompt, infer the task, and respond in the requested language. It handles translation, summarization, classification, question answering, data cleaning, and light coding across dozens of languages and scripts, and it can switch input and output languages in a single session. Checkpoints are available from small to very large so teams can trade accuracy for speed and footprint, and the open licensing makes it practical for academic work, benchmarking, and bespoke fine-tuning. In production it is a solid fit for multilingual assistants, localization tools, and retrieval-augmented apps where transparency and self-hosting matter. Limits include shorter context than modern models, weaker long-form reasoning, and the need for guardrails, but for multilingual instruction following with open weights, BLOOMZ remains a dependable baseline.