Modelpublic

yaoandy107/greyscope-v2-qwen3.5-4b-mlx-4bit

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes305downloads

Model Card

Greyscope v2 MLX 4-bit

This is the recommended Greyscope build for Apple Silicon. It is 2.4 GB and returns the same continuous ai_involvement score and human / AI-edited / AI-generated labels as the bf16 model.

Quick start

bash

git clone https://github.com/yaoandy107/greyscope
cd greyscope
uv sync --extra mac
uv run greyscope "Paste a paragraph here."

Mac performance

Measured on an M1 Pro with 32 GB unified memory and batch size 1:

Build	Peak memory	Load time	128 tokens	512 tokens	1,024 tokens
MLX 4-bit	3.0 GB	2.0 s	0.66 s	2.43 s	5.13 s
Transformers bf16	9.4 GB	16.4 s	0.74 s	2.54 s	5.01 s

Quantization check

MLX 4-bit matched bf16 quality across 3,007 external benchmark rows. See the release metrics for details.

Public evaluations, calibration, training details, and limitations are documented on the bf16 model card.