Granite-3.1-2b-instruct is a lightweight, instruction-tuned language model developed by IBM as part of the Granite 3.1 family. Designed for enterprise-grade performance, it is optimized for tasks such as summarization, retrieval-augmented generation (RAG), and function-calling. The model is released under the Apache 2.0 license, facilitating both research and commercial applications.
Built on a decoder-only dense transformer architecture, the model features approximately 2.5 billion parameters. A primary enhancement in the 3.1 version is the expansion of the context window to 128k tokens, up from the 4k limit of the previous 3.0 series. This extension enables the model to handle significantly larger documents and more complex multi-turn dialogues.
The model was trained on a massive dataset of 12 trillion tokens, comprising a mix of publicly available data with permissive licenses, synthetic data for long-context reasoning, and human-curated examples. It natively supports 12 languages, including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. For professional safety, IBM aligned the model using supervised fine-tuning and reinforcement learning to meet enterprise governance and risk standards.