Granite 4.2 3B

Granite 4.2 3B is a compact, dense reasoning language model developed by IBM as part of the Granite 4.2 family. It is specifically optimized for edge deployment and resource-constrained environments, providing advanced reasoning capabilities in a small-footprint package. The model is released under the Apache 2.0 license, emphasizing IBM's commitment to open-weight models for enterprise and academic use.

The defining feature of Granite 4.2 3B is its native reasoning (thinking) capability. Unlike previous generations, the model is trained to perform step-by-step chain-of-thought reasoning within specialized <think> tags before generating a final response. It supports flexible thinking modes—Full, Low-Effort, and Non-Thinking—allowing users to adjust the reasoning budget to balance between response depth and inference latency.

Technically, the model utilizes a dense decoder-only transformer architecture and supports a native context window of 128K tokens. It was pre-trained on a massive dataset of approximately 15 trillion tokens, followed by a multi-stage post-training process that includes supervised fine-tuning (SFT) and reinforcement learning (RL) focused on instruction following, safety, and reasoning trajectories.

In addition to reasoning, Granite 4.2 3B features native support for tool-calling and is capable of handling multilingual dialogues across 12 languages, including English, German, Spanish, French, and Japanese. While it is the smallest in its family, it retains high-precision XML-based tool calling and structured output capabilities, making it suitable for agentic workflows on local hardware.