IBM launches self-hosted Granite 4.2 LLMs with agentic training and reasoning focus

IBM released Granite 4.2, an open-weight family of large language models available in 3B, 8B, and 30B parameter sizes, designed for local deployment. The 8B and 30B variants include agentic reinforcement learning for tool use, while all models feature a 128,000-token context window and a reasoning-oriented approach via chain-of-thought. The release targets enterprises seeking predictable, cost-effective alternatives to frontier cloud models.

EXPANDED:

The Granite 4.2 lineup emphasizes practical deployment over headline-grabbing performance, with IBM positioning these models for organizations that prioritize predictable behavior and controlled costs over frontier-level capabilities. The agentic training applied to the 8B and 30B variants enables terminal interaction, web search, and external tool use, expanding what self-hosted systems can accomplish without cloud dependency.

Interest in local models has grown amid rising concerns about compute costs tied to cloud-based frontier systems, and this has spurred development of model routers that match prompts to appropriately sized models. Granite's open-weight approach also appeals to researchers and hobbyists