IBM Granite 4.2

The read

Granite 4.2 replaced the 3.x line in August with dense models at three, eight and thirty billion parameters, each able to switch reasoning on and off. They are Apache-licensed, documented with the care regulated buyers expect, and small enough to run on ordinary servers. The 30B model is genuinely useful on coding tasks, though none of them competes with the large open models on hard reasoning. Choose it for provenance and paperwork, not for headlines.

Where it shines

• Apache 2.0, with IBM's documentation behind it
• Reasoning can be switched on only when a task needs it
• Sizes from three to thirty billion parameters fit ordinary hardware

Worth knowing first

• Well behind the large open models on hard reasoning
• Text only, with a 128,000-token default context