Granite 3.0
IBM's open-weight enterprise model family for regulated industries
Granite 3.0 is IBM’s third-generation family of open language models, released on October 21, 2024 for enterprise use cases like customer service, cybersecurity, and RAG pipelines. The lineup includes dense 2B and 8B models trained on 12 trillion tokens plus 1B and 3B mixture-of-experts variants, all released under the Apache 2.0 license so they can be fine-tuned and deployed commercially without restriction. IBM positioned Granite 3.0 against similarly sized models from Meta and Mistral, reporting competitive or better scores on benchmarks covering reasoning, code, and math while keeping the models small enough to run cheaply on standard GPUs. The initial release shipped with a 4K context window, later extended to 128K tokens in the Granite 3.1 update released in December 2024. Granite 3.0 is available through Hugging Face, Ollama, and IBM’s watsonx platform, and it forms the base for IBM’s related Granite Guardian, Code, and Vision models.