IBM released Granite 4.2 on Tuesday, its first family of dense reasoning models, with the two largest sizes trained to act inside real software and web environments rather than just answer questions.
The family ships in 3B, 8B, and 30B parameter sizes under the Apache 2.0 license. All three get a switchable thinking mode: full chain-of-thought reasoning is the default, a non-thinking mode answers directly, and a low-effort setting spends a small reasoning budget on easy questions.
The training pipeline starts from scratch on roughly 15 trillion tokens with a five-phase schedule that stretches the context window to 512K tokens, followed by fine-tuning on about 7.2 million samples of chain-of-thought and agentic-trajectory data. The 8B and 30B models then move through three reinforcement-learning stages inside live environments, earning rewards only when tasks actually complete. A final RLHF pass applies a reasoning-length penalty to curb verbose chains.
Tool calling is built into the chat template using the OpenAI function-calling format, so the models plug into vLLM or SGLang harnesses without adapters. IBM evaluated the family on agentic coding, tool use, and long-context suites via its NeMo Evaluator SDK, with scores rising consistently with model size.
All weights are publicly downloadable, including the flagship 30B, whose model card confirms it was trained with asynchronous GRPO and separate GPU pools for generation and training.
