IBM continues to redefine the artificial intelligence landscape with the launch of its latest generation of Granite models at the annual TechXchange conference. This advance is part of its process of continuous innovation and expansion of its open source portfolio, published under the permissive Apache 2.0 license.

Granite 3.0: A varied and efficient offer

The Granite 3.0 model family consists of several variations suitable for various use cases, including:

• Four versions of general-purpose LLM
• Two models focused on security
• Four MoE-like models, optimized for efficient inference and reduced latency
• A Granite Time Series model, specially designed for time series

The dense versions of 2B and 8B parameters were rigorously trained on 12 trillion tokens. For their part, the MoE models of 1B and 3B used 10 trillion tokens, with 400M and 800M parameters activated respectively.

Adapt to all business needs

IBM presents Granite 3.0 models as key tools for enterprise AI, delivering optimal performance for varied tasks such as:

• RAG (Retrieval-Augmented Generation)
• Classification
• Synthesis
• Feature extraction
• Using built-in tools

These compact versions facilitate fine-tuning with company-specific data and integrate harmoniously into various work environments. MoE models, being lighter, are perfectly suited to applications requiring low latency and can be deployed on CPU infrastructures.

Economy and efficiency of models

By combining a small Granite model with enterprise data via the InstructLab technique, IBM enables businesses to achieve task-specific performance at a reduced cost, estimated to be between 3x and 23x less expensive than large “frontier” models. This represents a significant step forward for organizations seeking to optimize their investments in artificial intelligence.

Intellectual property and partnerships

IBM doesn’t stop there. To build customer confidence in using their data with Granite models, the firm offers an intellectual property indemnity on watsonx.ai. This mechanism encourages companies to share their data for training models.

Performance and precision of Granite 3.0

When it comes to performance, the Granite 3.0 8B Instruct model outperforms other comparable open source models, such as Meta’s Llama 3.1-8B and Mistral AI’s Mistral-7B, according to Hugging Face’s OpenLLM Leaderboard. On the AttaQ security benchmark, also developed by IBM, Granite 3.0 8B ranks first.

Granite Time Series models: A notable evolution

IBM also has their Granite Time Series models, having been pre-trained on three times as much data. These models show superior performance even compared to those ten times larger offered by competitors such as Alibaba and Google.

Accessibility and integration on platforms

Granite 3.0 models are now available for download on Hugging Face, while being integrated into the ecosystem of technology partners such as AWS, Docker, Domo, Qualcomm, Salesforce and SAP. The Instruct and Guardian variants of LLM Granite 3.0 (2B and 8B) are available for commercial use on the IBM watsonx platform.

Future advancements and problem solving

In the near future, IBM plans to add support for a 128K context window, as well as multimodal document understanding capabilities for its 3.0 models. At the same time, the company is innovating by developing autonomous AI agents, including advanced reasoning and multi-step problem solving capabilities, with updates planned in 2025.

With Granite 3.0, IBM asserts its leadership position in providing artificial intelligence solutions that combine performance, cost effectiveness and adaptability to specific business needs. The consultation between technological innovation and strategic partnership looks promising for the future of AI.