Granite 4.2 3B vs Granite 4.2 8B

These Granite 4.2 checkpoints share Apache licensing, text output, reasoning controls and documented native 128K context. Their main deployment difference is model size. The 3B checkpoint is a candidate for a smaller local assistant, while 8B provides a larger model to evaluate when a compact route misses task requirements. Neither size figure is a measured throughput result on your hardware.

Use identical prompts for extraction, a document-based answer and a short tool sequence. Include a case with missing evidence and one where the first tool result is incomplete. Check factual support, valid arguments and recovery behaviour rather than judging only writing style. Preserve the same thinking-history setting, context and output budget across trials.

For local serving, include memory, concurrent requests, preprocessing and the time to the first useful result. Test long-context extension separately from the native setup, since it changes the operational comparison.