IBM has released Granite 4.2, a family of dense reasoning language models with 3 billion, 8 billion and 30 billion parameters. The models are available under the Apache 2.0 license and support native tool calls plus thinking, non-thinking and low-effort reasoning modes.
All three were trained from scratch on roughly 15 trillion tokens through five phases, with the final phase extending the context window to 512,000 tokens. Supervised fine-tuning used about 7.2 million examples spanning agent tasks, software engineering, tool use, mathematics, multilingual work, science and safety.
The 8B and 30B versions received an additional reinforcement-learning stage in sandboxed environments. That training covered activities such as editing and running code, operating a terminal and searching the web. When served through an OpenAI-compatible endpoint, the models emit standard function-calling output, reducing integration work for existing agent frameworks.
Granite 4.2 is a technical model release rather than a guarantee of equal performance across all tasks. The three sizes share an architecture, but only the larger two received the dedicated agentic reinforcement-learning block, and real deployment requirements still depend on model size and chosen precision.