IBM releases Granite 4.2 with reasoning and agentic RL in three sizes

IBM has released Granite 4.2, a new family of dense reasoning language models in 3B, 8B and 30B parameter sizes. The company says all three are available under the Apache 2.0 license, support native tool calling, and can switch among thinking, non-thinking and lower-effort reasoning modes (IBM Granite).

The release matters less as another set of weights than as a detailed account of how IBM is trying to make smaller open models useful for agent workloads. The 8B and 30B variants received reinforcement learning in software-engineering, terminal and web-search environments; the 3B model did not receive that agentic training block (IBM Granite).

What changed

IBM says it pretrained each model from scratch on roughly 15 trillion tokens using five phases, with the final phase extending training to 512,000-token sequences. Supervised fine-tuning then used about 7.2 million samples, with agentic data making up 31.6% of the mixture. The company describes a multi-stage reinforcement-learning pipeline built around asynchronous Group Relative Policy Optimization, followed by reinforcement learning from human feedback (IBM Granite).

The capability split is important. All three models received foundational reinforcement learning and alignment. IBM says only the 8B and 30B models were trained to edit code, operate a terminal and search the web in live sandboxed environments, while the 3B model takes a shorter training path focused on foundational reasoning (IBM Granite).

For serving, IBM says Granite 4.2 emits tool calls in the OpenAI function-calling format when exposed through an OpenAI-compatible endpoint such as vLLM. The release also has an SGLang serving recipe. Those details reduce integration work for platform teams already using common open-model serving stacks, though they do not establish application-level reliability (IBM Granite).

What platform teams should check

The release post includes benchmark results and extensive training detail, but the performance claims are IBM's own reporting. Teams considering Granite 4.2 should test the exact model size, reasoning mode, quantization and tool harness they intend to operate rather than treating the family as interchangeable.

The practical choice is also not simply largest versus smallest. The 3B model offers the same license and tool-call interface, but lacks the agentic reinforcement-learning stages. The 8B and 30B models carry the fuller agent training recipe at higher serving cost. That makes workload-specific evaluation—especially tool-call correctness, multi-step recovery and latency—the useful next step.