Dynamic Beating AI News Flash: IBM has open-sourced the next-generation language model Granite 4.2, offering three tiers: 3B, 8B, and 30B. All three models include native inference and support options for full deliberation, low-effort deliberation, or direct answering, as well as tool invocation. The model weights are released under the Apache 2.0 license.
The key upgrade this time is the Agent. The 8B and 30B models have specifically incorporated Agent reinforcement learning, not just working on datasets but directly engaging in tasks within real code repositories, terminals, and web search environments. Writing code is based on whether tests pass, operating the terminal depends on whether tasks are truly completed, and then rewarding the model based on the results. The 3B model did not undergo this phase.
The standout performance comes from the smaller models. Artificial Analysis has awarded the Granite 4.2 3B model an Intelligence Index score of 14, ranking it 2nd among 46 similar open-weight models, with the median score among peers being only 4 points. The 8B model scored 20 points, also significantly higher than the peer median of 9 points.
In IBM's internal testing, the 30B model achieved 57% on SWE-Bench Verified and 89.17% on AIME25. In comparison to Granite 4.1, the most evident addition in this generation is not to continue piling on foundational capabilities but to formally enable small models for inference and to have the 8B and 30B models directly learn how to be Agents in real environments.
