IBM Granite 4.0 h-small

IBM Granite 4.0 H small is part of IBMs next-generation line of language models, featuring a hybrid architecture that merges Mamba-2 sequence state modeling with selective transformer layers. The model is a mixture-of-experts (MoE) variant with 32B total parameters and 9B active parameters, enabling dramatic improvements in inference and memory efficiency. This model is released under the Apache 2.0 license with cryptographic signing of model checkpoints. Granite 4.0 models are also ISO 42001 certified, making it one of the first open model families to satisfy international standards in trustworthy AI. Granite 4.0 H small is optimized for long-context and multi-session usage. All Granite 4.0 models have been trained on data samples up to 512K tokens in context length. Performance has been validated on tasks involving context length of up to 128K tokens, but theoretically, the context length can extend further. It is engineered to deliver over 70% reduction in memory usage relative to conventional transformer models in long-context or concurrent workloads. Typical use cases include compact deployments on consumer or enterprise GPUs, edge inference, local experimentation, multi-tool agents, and workloads requiring efficient memory footprint for long input spans. It can also serve as a building block within hybrid systems where its speed and efficiency are exploited for back-end tasks. Granite 4.0 H small demonstrates competitive performance relative to prior-generation models despite its much smaller active parameter footprint.