IBM Launches Granite 4.2: Open-Source Reasoning Models with Agentic RL

IBM has released Granite 4.2, its first family of dense, decoder-only reasoning language models, in three sizes: 3B, 8B, and 30B parameters. [2] Each model is pre-trained from scratch on roughly 15T tokens with a five-phase strategy that extends the context window to 512K tokens. [1] Unlike earlier Granite releases, which were instruction-following assistants, Granite 4.2 is built around explicit reasoning. [3]

Every model can emit a chain of thought before answering, and every model exposes a thinking / non-thinking switch plus a low-effort mode that spends a short reasoning budget on easy questions. [3] Supervised fine-tuning followed pre-training, utilizing about 7.2 million samples of chain-of-thought, reasoning, and agentic-trajectory data. [2]

The multi-stage reinforcement learning pipeline features separate training runs targeting specific capabilities. [2] All three sizes run foundational RL on verifiable rewards, such as math problems with checkable answers and code graded by hidden unit tests. [2]

Only the 8B and 30B models continue into the agentic block, a three-stage process where the model acts inside a live environment and receives rewards based on task completion. [2] In the software-engineering stage, driven by the OpenHands harness, the model edits real repositories and passes only if the hidden test suite passes. [2]

The release of Granite 4.2 positions IBM strongly in the enterprise AI space by offering models with native tool calling and explicit reasoning under an Apache 2.0 license. [1][3] The agentic RL approach for the 8B and 30B models, utilizing real environments like OpenHands and live shells, demonstrates a commitment to practical software engineering and DevOps automation applications. [2][3] What the sources don't address: How the reasoning performance and execution speed of Granite 4.2's low-effort mode compare to similar tiered reasoning features from competitors like OpenAI or Anthropic.