IBM Granite 3.0 2B Instruct is a lightweight, open-source large language model designed for enterprise-grade instruction-following tasks. As part of the third generation of IBM’s Granite family, this 2-billion parameter model is optimized for high-performance applications in resource-constrained environments, offering a balance between speed, cost-efficiency, and accuracy. It is built upon a dense decoder-only transformer architecture and is released under the Apache 2.0 license.
The model was developed through a multi-stage training process, starting from the Granite 3.0 2B Base model, which was pre-trained on 12 trillion tokens of natural language and code data. The instruction-tuned variant utilizes techniques such as supervised fine-tuning, model alignment using reinforcement learning, and model merging. This training pipeline incorporates a mix of permissive open-source datasets and internally generated synthetic data to enhance reasoning and task adherence.
Granite 3.0 2B Instruct is specifically engineered for enterprise use cases including retrieval-augmented generation (RAG), summarization, text classification, and entity extraction. It features capabilities for tool use and function calling, enabling it to serve as a core component in agentic workflows. Additionally, the model supports 12 natural languages, including English, German, Spanish, French, Japanese, and Chinese, making it suitable for a variety of multilingual applications.