Granite 3.1 - IBM's new language model
Granite 3.1 is IBM's next-generation language model, boasting powerful performance and extended context processing capabilities. The Granite 3.1 model family includes four different sizes and two architectures: a dense 2B model and an 8B parametric model...
What is Granite 3.1?
Granite 3.1 is IBM's next-generation language model, boasting powerful performance and extended context processing capabilities. The Granite 3.1 model family includes four different sizes and two architectures: dense 2B and 8B parameter models, trained using a total of 12 trillion tokens; and expert hybrid MoE models: sparse 1B and 3B MoE models with 400M and 800M activation parameters respectively, trained using a total of 10 trillion tokens. Granite 3.1 supports context lengths of up to 128K tokens, enabling it to handle complex text and tasks. Granite 3.1 introduces a new embedding model covering 12 languages, enhancing multilingual support. The Granite Guardian 3.1 model features function call illusion detection capabilities, improving control and observability of tool calls.
Main features of Granite 3.1
- Extended context windowThe context window is expanded to 128K tokens, enabling models to handle larger inputs, perform longer consecutive interactions, and incorporate more information into each output.
- A brand new embedding modelIntroducing a series of new retrieval-optimized Granite Embedding models, offering different parameter sizes from 30M to 278M, and supporting multilingual support for 12 different languages.
- Function call illusion detectionGuardian 3.1 adds the ability to detect tool call illusions in its 8B and 2B models, enhancing control and observability over agents that make tool calls.
Technical principles of Granite 3.1
- Large Language Models (LLM)Based on the construction of a large language model, it is trained through deep learning and natural language processing techniques to understand and generate natural language text.
- Transformer ArchitectureUse a transformer architecture to capture long-distance dependencies in text.
- Context WindowExpanding the context window to 128K tokens allows the model to consider more information simultaneously, handle longer texts, and perform more complex tasks.
- Multilingual supportThe trained model can recognize and understand multiple languages, supporting text processing in 12 different languages.
Granite 3.1 project address
- Project official website:ibm-granite-3-1
- GitHub repository:https://github.com/ibm-granite/granite-3.1
- HuggingFace model library:https://huggingface.co/collections/ibm-granite/granite-31
Application scenarios of Granite 3.1
- Customer service automationBuild chatbots and virtual assistants to automate customer inquiries and support services, improving response speed and efficiency.
- Content creation and editingIt helps content creators generate, edit, and optimize text content, including articles, reports, and marketing materials.
- Enterprise search and knowledge managementBased on semantic search and vector search technologies, improve the enterprise's internal knowledge retrieval system and help employees quickly find the information they need.
- Language translation and localizationTranslation services and localization projects for multinational corporations.
- Risk management and compliance checksIn the financial and legal fields, it helps analyze and understand large amounts of documents to ensure compliance and risk management.