IBM has released Granite 4.2, a new family of open language models available in three sizes — 3 billion, 8 billion, and 30 billion parameters — all released under the Apache 2.0 license, according to MarkTechPost. That license matters: it is one of the most permissive in software, meaning companies can download the models, modify them, and use them commercially without paying IBM or asking permission.
The headline feature is a switch. MarkTechPost reports that every model in the family exposes a "thinking / low-effort / non-thinking" control, letting the user decide how much internal deliberation the model does before answering. In practice, that is a dial between cost and care — quick replies for simple questions, slower and more deliberate reasoning for hard ones. Reasoning tokens are expensive to generate, so being able to turn them off is a real budget lever.
Every model also supports native tool calling, meaning it can invoke external software — a search function, a database query, a calculator — rather than trying to answer everything from memory.
The two larger models go further. According to MarkTechPost, the 8B and 30B versions pass through an additional "agentic RL" training stage — reinforcement learning that specifically teaches them to edit code, rather than merely produce it. IBM has also published a technical writeup on the Hugging Face blog detailing how the family was built.
Why it matters: enterprises that want capable reasoning and coding models without vendor lock-in now have a permissively licensed option they can run on their own hardware, with an explicit control over how much thinking — and spending — each query costs.