1. Strong Reasoning for an 8B Model
One of the biggest changes in Granite 4.2 is the training process.
All three models go through a base reinforcement-learning stage focused on math, coding, and reasoning.
The 8B and 30B versions get an extra agent-focused reinforcement-learning stage aimed at software engineering, terminal coding, and search-driven workflows.
The benchmark results are hard to ignore:
- AIME25: 86.67%
- GPQA: 64.14%
- MMLU-Pro: 74.04%
For an 8B model, those are strong numbers.
AIME25 stands out in particular. Getting that level of math reasoning from a model this size suggests IBM put real effort into reasoning training rather than simply feeding it more data.
2. Coding Beyond Basic Autocomplete
Granite 4.2 was trained on 1 trillion tokens of synthetic code generated through IBM’s CodeAlchemy pipeline.
It scored 47.67% on SWE-Bench Verified.
That is a solid result for an 8B model.
It is also designed for more than basic code completion. The model can work through codebases, handle multi-step development tasks, and run commands in terminal environments.
When paired with an agent framework such as OpenHands, it can plan a task, edit code, run commands, inspect the result, and continue from there.
That is much more useful in real development work than simply being good at writing a short function.
3. Tool Use With More Deliberation
Granite 4.2 8B has also been trained specifically for tool use.
Instead of immediately calling the first tool that looks relevant, it can reason about which tool to use and why before taking action.
It scored 68.05% on τ³-bench.
That benchmark is closer to a real multi-tool workflow than a single function call.
If an agent has to query a database, call an API, run a script, and combine the results, choosing the wrong tool early can ruin everything that follows.
So the important question is not just whether the model can call tools. It is whether it can choose the right one.
4. Three Reasoning Modes
Granite 4.2 8B lets you switch between three reasoning modes inside the same model:
- Full reasoning: for math, code, and harder tasks
- No reasoning: for simple requests where speed and lower token use matter more
- Low-effort reasoning: a middle ground for moderately difficult tasks
This is more useful than it may sound.
A basic classification request and a complicated code analysis do not need the same amount of reasoning.
Running both at maximum reasoning would just waste time and tokens.
Being able to adjust that without maintaining multiple models is a practical advantage.
What Changed From Granite 4.0?
Granite 4.2 changes the training approach more than the basic idea of the model.
All three sizes now use base reinforcement learning, while the 8B and 30B models get an additional agent-focused reinforcement-learning stage.
IBM also added the 1 trillion tokens of synthetic code training and a new intermediate training stage.
If you want the short version:
Granite 4.0 felt like a smart answerer. Granite 4.2 starts to feel more like something that can actually carry a task through to the end.
The focus is no longer just on getting the answer right. It is also about planning, using tools, executing steps, and staying on track.
Comments (0)