IBM Granite 3.0: open enterprise models

IBM has launched Granite 3.0, the latest iteration of its large language models (LLMs), designed for enterprise applications. This version emphasizes a balance of performance, safety, and cost-efficiency. The flagship model, Granite 3.0 8B Instruct, is an instruction-tuned LLM trained on over 12 trillion tokens across multiple languages, excelling in both academic and enterprise benchmarks. The models are open-source under the Apache 2.0 license, with detailed disclosures on training data and methodologies, reinforcing IBM's commitment to transparency and responsible AI. The Granite 3.0 suite includes various models tailored for different tasks, including cybersecurity and natural language processing. Notably, the introduction of Mixture of Experts (MoE) models enhances inference efficiency, while speculative decoding techniques significantly speed up text generation. Additionally, the Granite Guardian models provide advanced safety features to monitor and manage risks associated with LLM outputs. Future updates are planned to expand model capabilities, including increased context windows and multimodal functionalities. The models are available on the IBM watsonx platform and through various partners, emphasizing IBM's focus on sustainability by utilizing renewable energy for training.

- IBM Granite 3.0 features advanced LLMs optimized for enterprise use.

- The models are open-source, promoting transparency and responsible AI practices.

- New Mixture of Experts models enhance inference efficiency for low-latency applications.

- Speculative decoding techniques improve text generation speed significantly.

- Granite Guardian models offer comprehensive risk and harm detection capabilities.