The Granite-3.0-8B-Instruct is an 8-billion parameter language model developed by IBM as part of its third-generation Granite model family. Released under the Apache 2.0 license, it is designed as a foundational model for enterprise-grade applications, focusing on reliability, safety, and efficiency for business-centric workflows.
The model is built on a decoder-only dense transformer architecture that incorporates Grouped-Query Attention (GQA), Rotary Positional Embeddings (RoPE), and SwiGLU activation functions. It was trained on 12 trillion tokens of data encompassing 12 natural languages and 116 programming languages, utilizing a novel two-phase training method to optimize its performance for instruction-following tasks.
Specialized for core enterprise requirements, Granite-3.0-8B-Instruct excels in Retrieval-Augmented Generation (RAG), text summarization, entity extraction, and classification. It also features robust capabilities for function calling and tool use, allowing it to integrate into complex agentic workflows and AI assistant applications.
IBM emphasizes transparency and safety with this model by providing disclosures regarding training data and evaluation methodologies. The model's safety performance is benchmarked against dimensions such as social bias, harm, and unethical behavior, making it suitable for deployment in business environments where safety and trust are primary concerns.