IBM has introduced Granite 4.0, the latest iteration of its open-source large language models (LLMs) aimed at delivering high performance while reducing memory and cost demands. Founded in 1911, IBM remains a significant player in the tech industry and has garnered attention with Granite 4.0, which showcases notable performance on third-party benchmarks. The models are licensed under Apache 2.0, allowing developers to adapt and use them for commercial purposes. This development also positions the U.S. competitively against the growing capabilities of Chinese LLMs.
The Granite 4.0 release features a hybrid design that merges transformer and Mamba architectures. While transformers have been the preferred structure for LLMs since their introduction in 2017, they can be inefficient for lengthy text due to high computational and memory requirements. Mamba, a newer architecture developed in 2023, processes tokens sequentially instead of in a comprehensive manner, leading to more efficient handling of long documents and multiple requests.
Granite 4.0’s architecture allows for significant reductions in GPU memory use, reporting over a 70% decrease compared to traditional LLMs. Benchmarks indicate that the new models compete effectively with larger systems, excelling in tasks such as instruction-following and retrieval-augmented generation. Furthermore, Granite models are the first to achieve ISO/IEC 42001:2023 certification, which underscores international standards for accountability and data privacy.
IBM plans to continue expanding the Granite family, with additional models expected by the end of 2025. The models are currently available on platforms like Hugging Face and IBM’s watsonx.ai, with support expected on Amazon SageMaker and Microsoft Azure AI Foundry. By providing a robust and scalable solution for enterprises, IBM aims to address the need for efficient and legally compliant AI applications while signaling a competitive stance in the global AI landscape.
Source: https://venturebeat.com/ai/western-qwen-ibm-wows-with-granite-4-llm-launch-and-hybrid-mamba-transformer