IBM released Granite 4.2, the newest members of its open-weight Granite family designed to be downloaded and self-hosted.
The models ship in 3B, 8B, and 30B parameter variants and include a native 128,000-token context window to carry longer chains of thought.
IBM keeps a decoder-only approach and says the 8B and 30B variants were put through an agentic reinforcement-learning block to expand capabilities such as terminal use, web search and external-tool calling, while the 3B supports tools without the same specialised post-training.
The company described the release as “Granite 4.2 is the reasoning-focused release of the Granite language-model family,” highlighting its emphasis on chain-of-thought style reasoning.
Granite 4.2 continues a year-long shift by IBM toward compact, locally runnable models that can be deployed on cloud, on-premises or at the edge and distributed through open channels for enterprises and developers.
The launch dovetails with IBM’s broader push into enterprise AI infrastructure this year, including new managed platforms and data-streaming integrations intended to support agentic applications.
Reviewers and users should expect higher compute and slower responses for heavier reasoning workloads, a trade-off that has shaped the wider local-LLM market this year.
