IBM Unveils Granite 4.0: Open-Source LLM Family Combining Mamba-2 and Transformer
IBM has unveiled Granite 4.0, an open-source Large Language Model (LLM) family that combines the efficiency of Mamba-2 with the versatility of Transformer. This new family includes four models, each offering a unique balance between size and performance. The Granite 4.0 family consists of H-Small (32B), H-Tiny (7B), H-Micro (3B), and Micro (3B) models. Notably, Granite 4.0 reduces serving memory by over 70% for long-context and multi-session inference, making it highly efficient. IBM has licensed Granite 4.0 under Apache-2.0 and cryptographically signed it for security. The models are available in BF16 format with quantized and GGUF conversions for flexibility. They were trained on samples up to 512K tokens and evaluated up to 128K tokens. Granite 4.0's hybrid Mamba-2/Transformer stack and active-parameter Mixture-of-Experts (MoE) enable lower total cost of ownership (TCO) and easier production. IBM highlights its performance in instruction following and tool-use benchmarks like IFEval and BFCLv3. The models are available on platforms such as watsonx.ai, Docker Hub, and Hugging Face. IBM's Granite 4.0 is the first open model family with accredited ISO/IEC 42001 AI management system coverage. With its efficient serving memory, flexible licensing, and strong performance, Granite 4.0 is poised to make a significant impact in the open-source LLM landscape.
