Granite 4.0 H-Micro
Granite 4.0 H-Micro is an instruction model for text generation, extraction, coding and tool tasks. Its published parameter count is 3B dense hybrid. The release supports 12 languages.
Model design and training
H-Micro combines four attention layers with 36 Mamba-2 layers. Its hidden size is 2,048. The architecture uses no positional embeddings. It is post-trained from the matching base with supervised fine-tuning, reinforcement learning and model merging.Granite 4.0 H-Micro model card
Practical use
Practical guidance. Evaluate H-Micro when a small hybrid model fits your local assistant. Confirm that the serving engine supports its Mamba layers and expected cache behavior. Compare it with dense Micro using the same documents and output budget, since parameter count alone does not predict serving performance.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | ibm-granite/granite-4.0-h-micro |
| Context window | 131,072 tokens Published model context; deployment settings can impose a smaller limit. Granite 4.0 H-Micro model card |
| License | Apache 2.0Granite 4.0 H-Micro model card |
| Parameters | 3B dense hybridGranite 4.0 H-Micro model card |
| Architecture | Hybrid Mamba-2 and attentionGranite 4.0 H-Micro model card |
| Hosted access | watsonx: deploy on demand.IBM watsonx.ai foundation model library |
Related models
- Granite 4.0 H-Tiny
IBM · Granite 4.0
A hybrid instruction model with 7B total and 1B active parameters for text, code and tool tasks.
Launch date not confirmedAPI · Downloadable · Provider product - Granite 4.0 Micro
IBM · Granite 4.0
A 3B dense instruction model for text generation, extraction, coding and tool tasks.
Launch date not confirmedDownloadable