IBM's Granite 4.2 Adds Reasoning and Agentic Training to Small Open Models

IBM's Granite 4.2 is its first family of small open models built around explicit reasoning, released at 3B, 8B, and 30B parameters under an Apache 2.0 license. Each model can switch between a thinking mode that generates chain-of-thought and a faster non-thinking mode, so a caller can pay for step-by-step reasoning only when a task needs it. The architecture is conventional for this class: a decoder-only transformer with grouped query attention, rotary position embeddings, and SwiGLU feed-forward layers, extended to a 512K token context.

The more interesting part is the training pipeline that IBM describes in detail. Pre-training runs across five phases on roughly 15 trillion tokens with progressive quality filtering, followed by supervised fine-tuning on 7.2 million samples that mix agentic and non-agentic instruction data. Then comes a staged reinforcement-learning curriculum: foundational RL on verifiable tasks first, then agentic RL for the 8B and 30B models on real software engineering, terminal commands, and web search, closing with RLHF alignment. Training the models to act in real environments rather than simulated ones is what IBM credits for the coding results, with the 8B and 30B variants scoring 47.67 and 57.00 percent on SWE-Bench. All three sizes support native tool calling compatible with standard agentic frameworks.

Why it matters

If you build agents on open weights and cannot run a frontier model on your own hardware, a 30B model that scores 57 percent on SWE-Bench under a permissive license is a practical option to test. The documented RL recipe also gives teams a concrete template for adding agentic behavior to their own small models.