Z3NN001/gemma-4-21b-a4b-it-REAP-mlx-bfloat16

Gemma-4-21B-A4B-it-REAP (MLX bfloat16)

This repository contains the full-precision bfloat16 conversion of 0xSero/gemma-4-21b-a4b-it-REAP to the MLX format, specifically optimized for Apple Silicon (M1/M2/M3/M4).

Model Highlights

• Architecture: Gemma 4 Multimodal with Active Blocks (A4B).
• Optimization: REAP (Reasoning Enhancement and Active Pruning) for superior vision-language performance.
• Precision: bfloat16 (No quantization loss, identical to original weights).

Technical Details

The conversion was performed using the mlx-vlm library on local hardware to ensure bit-perfect compatibility with the MLX ecosystem.

• Hardware: Mac Mini (M-Series) with 32GB RAM.
• Software: mlx-vlm (latest version).
• Format: Native MLX safetensors.

Usage

You can run this model using the mlx-vlm library.

Installation

pip install mlx-vlm

Credits
Original Model: 0xSero/gemma-4-21b-a4b-it-REAP

MLX Conversion: Z3NN001