IBM's Granite 4.2: Three Sizes, One Big Leap Into AI Reasoning
IBM has released Granite 4.2, a family of reasoning-focused AI models that can think through hard problems step by step, use tools, and even browse the web, all for free under an open licence.
Key points
- IBM released Granite 4.2, a family of three AI models (3B, 8B, and 30B), in mid-2025 under a free Apache 2.0 open licence.
- Each model was trained on roughly 15 trillion tokens, the pieces of text and code an AI learns from, across five distinct training phases.
- The two larger models (8B and 30B) can act as AI agents, software that carries out multi-step tasks on its own, including running code, using a terminal, and searching the web.
- All three models include a thinking/non-thinking switch, letting users trade speed for depth depending on the task.
- The full technical walkthrough was published on Hugging Face, where the models are also freely available to download.
IBM quietly dropped something worth paying attention to. Granite 4.2 is a new family of AI language models, the kind of software that understands and generates text, built specifically to reason through problems rather than just answer them.
There are three versions: a compact 3B, a mid-sized 8B, and a larger 30B. The number refers roughly to how many internal connections the model has. Bigger generally means sharper, though also heavier on computing power.
What does "reasoning" actually mean here?
Think of it like showing your working. A standard chatbot gives you an answer. A reasoning model works through the problem first, checks its logic, then replies. Granite 4.2 lets you flip between a "thinking" mode for tricky questions and a faster "non-thinking" mode when you just need a quick answer. There is also a low-effort middle setting that spends only a short reasoning budget on straightforward questions.
Say you paste in a complicated work contract and ask whether there is a penalty clause. In thinking mode, the model reads carefully, reasons about what it finds, and gives you a considered answer. In non-thinking mode, it is quicker but shallower. You choose.
What can the bigger models do that the small one cannot?
The 8B and 30B models go a step further: they have been trained as AI agents. That means they can take actions, not just produce text. Give one a coding task and it can write the code, run it in a sandboxed (contained, isolated) test environment, check whether it worked, and fix it if it did not. It can also search the web or operate a command-line terminal.
IBM trained the models on roughly 7.2 million example tasks, including software engineering, tool use, maths, and multilingual text, then ran them through a reinforcement learning pipeline (a training technique where the model learns by trial, error, and reward, much like a dog learning tricks). The 8B and 30B models did extra rounds of this inside real simulated environments.
How much does it cost, and what is the privacy situation?
The models are released under the Apache 2.0 licence, which means free to download, use, and modify, including for commercial projects. You can run them on your own hardware, which means your data stays on your machine rather than passing through IBM's servers.
That said, running the 30B model locally demands serious computing power. Most people will access these through a cloud provider, in which case the usual caution applies: check what the provider does with your inputs before you paste anything sensitive.
| Model | Parameters | Context window | Agent skills |
|---|---|---|---|
| Granite 4.2 3B | 3 billion | 512K tokens | Tool calling |
| Granite 4.2 8B | 8 billion | 512K tokens | Tool calling, code, web, terminal |
| Granite 4.2 30B | 30 billion | 512K tokens | Tool calling, code, web, terminal |
All three models support a 512K token context window. A token is roughly three-quarters of a word, so 512K tokens is enough to feed in a very long document (think a full novel) in one go.
Common questions
Can a normal person actually use these models?
Yes, though not without some setup. The models are free to download from Hugging Face and IBM's GitHub, but running them yourself requires a reasonably powerful computer. Most non-technical users will find them through cloud platforms that host them with a simple interface.
Is this better than ChatGPT or Claude?
Granite 4.2 is competitive with models of similar size, and the open licence is a genuine advantage for businesses and developers who want control over their data. For casual home use, the big consumer chatbots are still easier to get started with.
