IBM Unveils Granite 4.2 Reasoning Models Featuring Agentic RL and 512K Context

IBM's Granite 4.2 is a family of dense, decoder-only reasoning language models available in 3B, 8B, and 30B sizes. [1] Each model is pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy that extends the context window to 512K tokens. [1] The models are released under the open-source Apache 2.0 license. [1]

The models undergo supervised fine-tuning on chain-of-thought, reasoning, and agentic-trajectory data, followed by multi-stage reinforcement learning. [1] The 8B and 30B models specifically undergo an agentic reinforcement learning process that teaches them to operate as agents inside real sandboxed environments, which includes calling tools, editing code, driving a terminal, and searching the web. [1] Every model in the family features native tool calling, emitting calls in the OpenAI function-calling format, and includes configurable thinking and non-thinking modes depending on the task's required deliberation. [1]

The release of Granite 4.2 positions IBM directly in the growing open-source reasoning model space, offering scalable 512K context options under a permissive Apache 2.0 license. By embedding agentic RL directly into the post-training pipeline for the 8B and 30B models, IBM addresses the increasing demand for models that can independently navigate sandboxed environments, execute code, and call tools. What the sources don't address: How Granite 4.2's reasoning and agentic capabilities benchmark against comparable open-weight models.