Granite 4.0 H Small
IBM · 2025-10-02 · 32.2B parameters
Granite 4.0 H Small is IBM's long-context instruction-tuned Granite 4.0 model, with 32,207,337,984 checkpoint parameters and a hybrid architecture comprising 36 Mamba-2 layers, four grouped-query-attention layers, and fine-grained mixture-of-experts feed-forward blocks; IBM reports approximately 9 billion active parameters. IBM positions it for enterprise agentic workflows such as multi-tool agents and RAG and reports over 70% lower RAM requirements for long inputs and concurrent batches than conventional Transformer models. Its weights are released under Apache 2.0, and it belongs to the Granite language-model family covered by IBM's ISO/IEC 42001 certification.