IBM has launched Granite 3.0, an open-source family of AI models designed for enterprise use, which allows for flexible deployment and enhances capabilities by leveraging proprietary data. The introduction of Instruct Lab enables enterprises to enrich the Granite models with their own data efficiently, while IBM continues to explore advancements in quantum computing and its integration with AI.
In a recent visit to IBM’s New York office, the speaker was introduced to two exciting AI projects, particularly focusing on the launch of Granite 3.0, an open-source family of models. These models are designed for enterprise use and come with an Apache 2.0 license, allowing for flexibility in deployment. Granite 3.0 includes various sizes, including a very small model that can run efficiently on devices. The models are built to be general-purpose, enabling users to supplement them with external knowledge, which is crucial given that much of the publicly available data has already been utilized in existing models.
Granite 3.0 aims to address the challenge of accessing valuable data that is often behind paywalls or authentication barriers. IBM’s approach allows enterprises to leverage their proprietary data, enhancing the model’s capabilities in specific verticals. The Granite models are available in two main flavors: general-purpose and a mixture of experts, which optimizes performance by activating only a portion of the model’s parameters during inference. This makes them suitable for deployment in various environments, including on-device applications and low-latency scenarios.
The speaker highlighted the performance of Granite 3.0 in comparison to other models, showcasing its superiority in tasks such as retrieval-augmented generation, classification, summarization, and tool use. Benchmarks indicate that Granite 3.0 outperforms similar-sized models in various categories, making it a strong contender for enterprise applications. Additionally, IBM has introduced the Guardian family of models, which focuses on alignment to ensure that the models adhere to best practices and guidelines in business environments.
Another significant development discussed was Instruct Lab, a new technique that allows enterprises to enhance the Granite model with their own data without overwriting the original knowledge. This method serves as a middle ground between traditional fine-tuning and retrieval-augmented generation, enabling users to add external knowledge efficiently. Instruct Lab is also open-source, allowing communities to collaborate on improving models without the need for extensive retraining.
Lastly, the speaker touched on IBM’s advancements in quantum computing, noting the company’s long-standing commitment to this field. The integration of AI and quantum computing was highlighted as an area of interest, with potential implications for future developments. The video concludes with an invitation for viewers to explore the new Granite models and Instruct Lab, emphasizing the open-source nature of these projects and encouraging engagement from the audience.