IBM open-sources Granite 4.2: Major inference upgrade, Mamba support removed.

Beating AI Express: IBM has open-sourced its new-generation Granite 4.2 language model, available in three variants: 3B, 8B, and 30B. All three models feature native reasoning, capable of switching between full reasoning, low-intensity reasoning, and no reasoning, as well as direct tool calling. The 8B and 30B versions additionally come with dedicated agent reinforcement learning, enabling them to execute tasks in real code repositories, terminals, and web search environments. Training rewards are directly tied to task outcomes, such as whether code passes tests or terminal tasks are completed. All three models are released under the Apache 2.0 license. Independent evaluations show the small models performed surprisingly well: Artificial Analysis awarded the 3B version 14 points, ranking it 2nd among 46 open-weight models of the same parameter size; the 8B version scored 20 points, significantly higher than the median of 9 for its peer group. In IBM’s internal tests, the 30B version achieved 57% on SWE-Bench Verified and 89.17% on AIME25. There is an architectural shift: Granite 4.0 featured a Mamba-2/Transformer hybrid architecture, which IBM claimed reduced memory requirements by over 70% in long-context and multi-concurrency scenarios; Granite 4.2 has switched back to a pure dense Transformer architecture. Compared to the previous generation’s focus on inference efficiency, this generation clearly prioritizes inference and agent capabilities.