Granite 4.0 H-Small
Granite 4.0 H-Small is an instruction model for text generation, extraction, coding and tool tasks. Its published parameter count is 32B total; 9B active. The release supports 12 languages.
Model design and training
H-Small combines four attention layers with 36 Mamba-2 layers. Its mixture-of-experts design has 72 experts, with ten active experts in the architecture table. The hidden size is 4,096. Post-training targets instructions, retrieval and tools.Granite 4.0 H-Small model card
Practical use
Practical guidance. Consider H-Small for a shared service that needs more capacity than the compact Granite models. Its active parameter count does not mean it has the memory footprint of a 9B dense model. Compare throughput, latency and complete answer quality under your expected concurrency.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | ibm-granite/granite-4.0-h-small |
| Context window | 131,072 tokens Published model context; deployment settings can impose a smaller limit. Granite 4.0 H-Small model card |
| License | Apache 2.0Granite 4.0 H-Small model card |
| Parameters | 32B total; 9B activeGranite 4.0 H-Small model card |
| Architecture | Hybrid Mamba-2 and attentionGranite 4.0 H-Small model card |
| Hosted access | watsonx: pay as you go or deploy on demand.IBM watsonx.ai foundation model library |
Related models
- Granite 4.0 H-Micro
IBM · Granite 4.0
A 3B dense hybrid instruction model for text, code and tool tasks across 12 languages.
Launch date not confirmedAPI · Downloadable · Provider product