---
license: apache-2.0
library_name: mlx
tags:
- language
- granite-4.0
- mlx
base_model: ibm-granite/granite-4.0-h-tiny
pipeline_tag: text-generation
---

An MLX variant of IBM's Granite 4.0 H Tiny, a Mixture of Experts model which consists of 7 billion total parameters and 1 billion active. The raw model weights were converted into a 3 bit MLX format. The model averages 40tok/s at 3GB of RAM usage on a MacBook Pro (base M1, 8GB of unified RAM, 256GB of internal storage), making it ideal for lower-end devices.

This model may be unstable and sometimes spit out nonsense, be aware of that!