According to CoinMeta, IBM has released a new generation of language model, Granite 4.2, which comes in three versions: 3b, 8b, and 30b. All versions include native inference capabilities, allowing for switching between full thinking, low-intensity thinking, and no thinking at all, and they also support direct invocation of tools. The 8b and 30b versions have undergone specialized agent reinforcement learning, enabling them to perform tasks in real code repositories, terminals, and web search environments. Training rewards are based directly on task outcomes, such as whether the code passes tests or whether terminal tasks are completed. All three models utilize the Apache 2.0 license. In independent evaluations, the 3b version scored 14 out of 46, ranking second among similar open-weight models, while the 8b version scored 20, significantly higher than the median score of 9. In self-tests, the 30b version achieved a swe-bench of 57% and an verified of 89.17%. In terms of architecture, Granite 4.2 has shifted from a previous hybrid architecture of mamba-2 / transformer to a pure dense transformer architecture, with a focus on enhancing inference and agent capabilities.