IBM unveils Granite 4.2 family of open-weight language models
IBM has introduced the Granite 4.2 series, a set of open-weight large language models that can be downloaded and run locally.
IBM announced the Granite 4.2 family, expanding its portfolio of open-weight large language models that users can host on their own hardware. The series offers three size options and continues the decoder-only design of earlier versions. Each model now handles an extended context window, and the two larger sizes receive additional training for agentic reinforcement learning, allowing them to interact with terminals, browse the web, and employ external tools.
The smallest model also supports tool use, though without the specialized reinforcement learning. IBM emphasizes that Granite 4.2 targets functional reasoning, employing chain-of-thought techniques to carry intermediate results across multiple steps.
Why it matters
The launch gives developers broader access to powerful, self-hosted AI models with enhanced reasoning capabilities.
In this story
