Granite 4.2

Granite 4.2 is the latest release in IBM Research's language model series. IBM built it in three sizes—3B, 8B, and 30B—and released all of them for free. The number indicates how large the model is: bigger numbers mean the model can learn and hold more, but also need more computing power to run.

Up through earlier versions, Granite behaved like an assistant that faithfully carried out whatever it was told. Starting with 4.2, though, it has picked up the habit of organizing its thoughts before giving an answer. It's similar to writing out each step of a math problem instead of jumping straight to the final number. All three sizes share this thinking habit, but only the 8B and 30B were trained to actually take action—editing code, operating a terminal, and so on. The smallest model, the 3B, skipped that training.

The two larger models built up real tool-use skills by practicing, in order, fixing software, working with command-line tools, and searching the web. So even though they all carry the same Granite 4.2 name, the 3B is a model that's good at conversation only, while the 8B and 30B can both converse and handle practical tasks.

It shows up in headlines like "IBM's Granite 4.2 taught tool use only to the 8B and 30B." The key point is that models sharing the same name can have different real-world capabilities depending on size, so it would be a mistake to assume "if it's Granite 4.2, it can fix code just the same across all sizes." The 3B can think, but it was never trained to use tools.

If you can run Granite 4.2 8B or 30B, try asking it to solve a math problem or fix a code bug. Watch whether it writes out its reasoning before giving the final answer—that's the new reasoning habit at work. It's also worth asking the same question to the 3B and comparing how it answers without tool use.