Granite 4.0

model
IBM

Novel hybrid Mamba-2/Transformer + MoE architecture achieving 70% less memory and 2x inference speed vs. dense Transformers. H-Small: 32B total / 9B active. H-Tiny: 7B / 1B active. Micro: 3B dense. Nano: 350M–1.5B (runs in browser). First open models with ISO 42001 certification.

AA Intelligence Index: 23 (H-Small), 16 (Micro), 14 (H-1B Nano). Apache 2.0.

AnnouncementArtificial AnalysisOpenRouterHuggingFace

Model Details

Architecture MOE
Parameters 32B
Active params 9B

Variants

Name Parameters Notes
Granite 4.0 H-Small 32B —
Granite 4.0 H-Tiny 7B —
Granite 4.0 Micro 3B Dense architecture
Granite 4.0 Nano 1.5B Browser/on-device
open-weightmoeenterprise

Related