IBM recently unveiled its latest and most powerful line of AI models, known as Granite 3.0, at its TechXchange event. This new series of models aims to deliver optimized performance for a range of enterprise applications, as well as improve safety and efficiency in AI-driven tasks.

Granite 3.0 Lineup

1. General Purpose/Language Models

The Granite 3.0 models come in two sizes for general purpose use:

• 8B and 2B variations, featuring both Instruct and Base options.
• These models are tailored for tasks such as Retrieval Augmented Generation (RAG), summarization, classification, and entity extraction.
  More about Granite 3.0 models

2. Safety: Guardian Models

• The Guardian models also come in 8B and 2B sizes, focusing on adding safety guardrails.
• These models are designed to enhance safety during interactions by analyzing both user input and language model output for potential risks.
  Explore the Guardian models

3. Mixture-of-Experts

This portfolio focuses on diverse deployment needs and is specialized for handling complex and varied applications.

• It provides adaptability for enterprises and developers, enabling efficient resource allocation.
  More about Mixture-of-Experts models

Key Features and Capabilities of Granite 3.0

Performance and Flexibility

• Granite 3.0 models are claimed to be on par with, or exceed, similarly sized models from other leading AI providers.
• They offer high performance, flexibility, and control for businesses, available under the Apache 2.0 open-source license.
  Read more on Apache 2.0 licensing

Cost Efficiency for Enterprises

• By integrating smaller Granite models with internal enterprise data, organizations can achieve performance comparable to larger models while reducing costs by up to 23 times in some cases.
  Learn more about cost-saving integration

Transparency and Safety

• IBM emphasizes AI transparency and safety through a detailed technical report and a responsible use guide for Granite 3.0, outlining the data sets and benchmarking processes.
• Intellectual property indemnity is also offered on IBM’s watsonx.ai platform, ensuring business users receive additional legal protection.
  Read the technical report

Granite 3.0’s Breakthrough Performance

The 8B Instruct model has surpassed similar-sized models from Meta and Mistral in education metrics and ranks highest in safety according to IBM’s AttaQ safety benchmark.
Learn more about AttaQ benchmarks

Guardian 3.0: Enhanced Safety Features

• The Granite Guardian 3.0 models provide specialized safety guardrails that analyze both user input and model output, particularly for RAG-related challenges like groundedness and relevance.
  Explore the Granite Guardian 3.0 models

Accessing Granite 3.0 Models

• All Granite 3.0 models are available for download on Hugging Face.
  Visit Hugging Face for Granite models
• Commercial access options are provided through IBM’s watsonx platform, and IBM has partnered with other ecosystem participants to integrate these models into broader enterprise solutions.
  Access via watsonx

Future AI Advancements from IBM

Looking ahead, IBM will continue to expand its AI portfolio with a focus on autonomous AI agents to tackle increasingly complex challenges. Notable upcoming developments include new AI agent capabilities for IBM watsonx Orchestrate and broader AI agent features across its entire portfolio by 2025.
Read more about watsonx Orchestrate