Granite 4.2 brings native reasoning to enterprise agents

Details

IBM has released Granite 4.2, an updated family of open-weight language models in three sizes (3B, 8B, and 30B parameters), designed specifically for enterprise agentic workflows. The standout addition is native reasoning — built-in step-by-step 'thinking' that helps models plan before acting, evaluate trade-offs, and self-correct, rather than executing instructions blindly. The training pipeline was rebuilt on top of Granite 4.0 base models using multi-stage reinforcement learning. A 'foundational RL' phase sharpened math, science, coding, and tool-calling across all sizes. For the 8B and 30B models, a second 'agentic RL' phase then targeted enterprise-specific tasks: software engineering, terminal-based coding, and search-driven workflows. This approach combines verifiable rewards with reward-model-based evaluation. Practically, Granite 4.2 can navigate codebases, handle multi-step dev tasks, choose tools in the correct sequence, and verify results. IBM highlights the deployment flexibility — cloud, on-prem, and edge environments — with smaller sizes suited for high-throughput pipelines and larger sizes for complex reasoning workflows. Apache 2.0 licensing removes legal friction for enterprise adoption: organizations can download, fine-tune, and ship to production without restrictions. This is a credible competitive move against both commercial models and other open-weight alternatives like Llama and Qwen in the enterprise segment.