IBM has released Granite 4.2, a new family of reasoning-focused language models available in three sizes—3B, 8B, and 30B parameters. The models represent a significant advance in enterprise AI, built through a comprehensive training pipeline that includes 15 trillion tokens of pre-training, supervised fine-tuning on reasoning and agentic data, and a multi-stage reinforcement learning approach. All three models support native tool calling via an OpenAI-compatible API and feature a thinking/non-thinking switch that allows users to control reasoning depth based on task complexity.
The distinguishing feature of Granite 4.2 is its agentic reinforcement learning component, where the larger 8B and 30B models learn to operate as autonomous agents within real sandboxed environments. During this phase, the models acquire skills in tool calling, code editing and execution, terminal operations, and web search. The training data combines 31.6% agentic content—drawn from software engineering, tool use, and interactive tasks—with 68.4% traditional instruction-following data, encompassing approximately 7.2 million samples across diverse domains.
The Granite 4.2 models are released under the Apache 2.0 license, making them freely available for enterprise deployment. IBM's technical approach emphasizes both instruction-following and reasoning capabilities, positioning the models for use cases requiring chain-of-thought problem-solving and interactive tool use in production environments.
Key Points
IBM released three Granite 4.2 reasoning models (3B, 8B, 30B) built from Granite-4.1 base models pre-trained on 15 trillion tokens with 512K context windows
Unique agentic RL training teaches larger models to operate as agents, calling tools and executing code in real sandboxed environments
All models feature native tool calling via OpenAI-compatible API, thinking/non-thinking mode, and low-effort reasoning for simpler tasks
Training combines 31.6% agentic data from software engineering and interactive tasks with 68.4% traditional instruction-following data across 7.2 million samples
Models released under Apache 2.0 license with support in SGLang and compatibility with vLLM serving frameworks