Granite 4.2 3B
Granite 4.2 3B is a dense text model post-trained from Granite 4.1 3B Base. It supports full thinking, direct answers and a low-effort reasoning mode. IBM released it on 25 August 2026.
Model design and training
The training pipeline combines supervised fine-tuning, alignment and reinforcement learning. It uses public instruction data, synthetic tasks, agent traces and human-authored examples. The chat template controls thinking behavior.Granite 4.2 3B model card
Practical use
Practical guidance. Use the 3B model for local assistants with limited accelerator memory. Compare direct answers with low-effort reasoning on short extraction and coding tasks. Extra reasoning tokens can dominate response time even when the weights are small. Measure complete task success, not the length of the explanation.
Context and serving limits
The card advertises a 512K extension, but its configuration and serving examples use 131,072 tokens. Treat extended context as a separate deployment setting to verify. The 512K claim is not used here as the default endpoint limit.Granite 4.2 3B model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-4.2-3b |
| Context window | 131,072 tokens Default published configuration and architecture table. IBM also describes extension to 512K. Granite 4.2 3B model card |
| License | Apache 2.0Granite 4.2 3B model card |
| Parameters | 3BGranite 4.2 3B model card |
| Supported languages | 12 languages, including English, German, French, Chinese and Japanese.Granite 4.2 3B model card |
Related models
- Granite 4.1 3B
IBM · Granite 4.1
An instruction-tuned 3B text model for retrieval-based answers, extraction, coding and tool use.
Launch date API · Downloadable · Provider product - Granite 4.1 3B Base
IBM · Granite 4.1
A pretrained 3B dense decoder for text generation, further training and fill-in-the-middle code completion.
Launch date Downloadable