Granite is IBM’s family of open large language models aimed at enterprise software development and business workflows. A landmark release came on May 6, 2024, when IBM open-sourced its Granite code models under the Apache 2.0 license - a family of four sizes from 3 to 34 billion parameters, in both base and instruction-tuned variants. The code models were trained on roughly 3 to 4 trillion tokens spanning 116 programming languages, plus an additional 500 billion tokens of mixed code and natural language, and were designed for code generation, fixing, and explanation across languages like Python, JavaScript, Java, Go, C++, and Rust.
The permissive license was central to the pitch: IBM emphasized releasing high-performing, cost-efficient code models without usage restrictions, contrasting with the conditional licenses on some other open-weight releases. IBM has since extended Granite into general-purpose dense and mixture-of-experts language models and other modalities, distributed through Hugging Face, GitHub, IBM’s watsonx.ai, and Red Hat’s RHEL AI. Specific members and sizes beyond the original code models reflect IBM’s announcements as of the verification date.
Why business readers should care: Granite is a major vendor’s bet on genuinely open, business-licensed models built for code and enterprise tasks, giving regulated organizations an option they can self-host and audit without the licensing limits attached to some rival open releases.