IBM

Granite 4.0 H Micro 3B

3,2B

Parameters

1,8 Go

Q4_K_M file

≥ 3,3 Go

Recommended VRAM

131 072

Max context

Not measured yet

No card has run the TokenBench protocol on this model yet. The current protocol covers a selection of models; benchmarking any catalog model from the app is coming soon.

Measure my card →

Which cards should it fit on?

Cards on the site whose video memory covers the recommended 3,3 Go (file + 4,000-token context). An estimate, not a measurement. Click a card for the detailed estimate.

Technical sheet

Published: 2025-09-16
Architecture: 40 layers · 8 KV heads · dimension 64
Attention: Hybrid: 4 of 40 layers use standard attention, the others a fixed-memory mechanism (Mamba, linear attention) — a very light cache.
Context cache (f16): ≈ 8 MB per 1,000 tokens · 0,3 Go for 32,768 tokens
GGUF file: bartowski/ibm-granite_granite-4.0-h-micro-GGUF · ibm-granite_granite-4.0-h-micro-Q4_K_M.gguf Download (1,8 Go)
Official repo: ibm-granite/granite-4.0-h-micro

Public Hugging Face data (official config.json, GGUF repo). The context cache only counts full-attention layers.