IBM just put a new family of open-weight reasoners on the table, and I am glad they did. Granite 4.2 arrived on August 25 in three dense sizes, 3B, 8B, and 30B, all under Apache 2.0. If you have been looking for a local model that can think a step through before it calls a tool, this is the one I would sit down with this week.
What actually shipped
Granite 4.2 is a dense, decoder-only language family with switchable thinking, non-thinking, and low-effort modes. That toggle is the part I want students to feel in their hands. Run the same prompt with thinking on, then turn it off, and you can see where the extra planning helps and where it just costs tokens.
The Hugging Face card for the 3B lists 128K tokens of native context, with a claimed stretch to 512K. All three sizes went through foundational reinforcement learning covering math, science, coding, reasoning, and tool calling. The 8B and 30B got an extra agentic RL stage aimed at software engineering, terminal work, and search-driven tasks. IBM also describes a trillion tokens of synthetic code from their CodeAlchemy pipeline, plus a speculative decoding layer meant to speed inference without asking you to change how you call the model.
Apache 2.0 is the other reason this release belongs on a course syllabus. You can download the weights, fine-tune them, and put them in a student repo without a special conversation with legal.
A speech model from the same day
IBM also released Granite Speech 5.0 Turbo CTC, a 470 million parameter speech recognizer with no language-model backbone. That is a different tool sitting next to the reasoners: automatic speech recognition for laptops and phones, built to map audio to text quickly. IBM reports about 12,600 RTFx on a single H200, against about 6,000 for the current Hugging Face Open ASR leaders. Those are vendor numbers. If speech is your problem, reproduce them. If you came for the 3B reasoner, just know the speech SKU exists and is tiny on purpose.
There is a non-commercial variant trained on restricted-use data. Read the card before you ship anything.
How I would load them
The 3B is the one I would put on a laptop first. It is small enough to iterate with, and the thinking mode gives you a real experiment: same prompt, two settings, a notebook of where the plan holds up.
The 8B is the workstation size. That is where the extra agentic training starts to matter, especially if you want the model to pick tools in sequence and check its own work in a terminal.
The 30B is for when you have the VRAM and a tool-using prompt you actually care about. IBM positions it for deeper reasoning and heavier coding workflows. I would not start there. I would graduate to it once the 8B has shown you the shape of the task.
Early benches are still young. Some of IBM’s own Tau figures already differ between the research post and the Hugging Face card. That is normal on day two of a release. Download a checkpoint. Point it at a prompt from your own work. Write down where it plans well and where it still guesses.
Why the training story matters
IBM is pretty explicit that these agentic habits did not come from scale alone. They redesigned the recipe: supervised fine-tuning, then foundational RL on every size, then agentic RL on the two larger ones, plus RLHF alignment. The mid-training step is there to unlock more reasoning. The speculative decoding layer is there so serving more users does not feel like a tax.
For enterprise work that is the honest pitch. Tasks are ambiguous. They take many steps. A model that can retrieve the right information, choose tools, act in order, and verify the result is more useful than a model that only writes a fluent paragraph. Whether Granite 4.2 does that on your repo is the thing to test, not the thing to take on faith.
Time to dig deeper
Start with IBM’s Granite 4.2 announcement and the 3B model card. Load the 3B on a machine you already own. Turn thinking on and off. Then try the same tool-calling prompt on the 8B if you have the memory. Keep notes. That afternoon of tinkering will tell you more about this family than another launch-day screenshot.