IBM has open-sourced Granite 4.2, its latest lineup of language models, in three sizes: 3B, 8B, and 30B. All three pick up native reasoning, with options for full thinking, low-effort thinking, or just answering straight — and all three support tool calling. The model weights are released under Apache 2.0.
The bigger headline this time is agents. The 8B and 30B models got agent-specific reinforcement learning. Instead of only grinding through a dataset, they’re turned loose in actual code repositories, terminals, and web search environments. Write code? Success is whether the tests pass. Work in a terminal? Success is whether the job actually gets done. The model gets rewarded based on those outcomes. The 3B skips that training phase.
The smallest model, though, is the independent-benchmark standout. Artificial Analysis gives Granite 4.2 3B an Intelligence Index of 14 — 2nd place among 46 open-weight models in its class, where the median score is just 4. The 8B lands at 20, also comfortably above its class median of 9.
In IBM’s own testing, the 30B hits 57% on SWE-Bench Verified and 89.17% on AIME25. Compared with Granite 4.1, this generation isn’t really about piling on more raw capability. It’s about giving small models actual reasoning and letting the 8B and 30B learn to be agents in the real world.