North-Micro-Vision-Instruct-mxfp8

North Micro Vision Instruct — MXFP8 (MLX)

This repository contains an Apple MLX conversion of CohereLabs/North-Micro-Vision-Instruct. MXFP8 MLX quantization with group size 32.

It belongs to the North Vision MLX collection, which includes BF16, affine 4/5/6/8-bit, MXFP4, MXFP8, and NVFP4 variants.

Conversion details

• Source: CohereLabs/North-Micro-Vision-Instruct
• Format: MLX / MLX-VLM
• Quantization: bits: 8; group size: 32; mode: mxfp8
• MLX-VLM source revision: main at 7ee8eba3

The repository was regenerated and uploaded directly with the MLX-VLM conversion CLI:

python -m mlx_vlm convert \
  --hf-path CohereLabs/North-Micro-Vision-Instruct \
  --mlx-path North-Micro-Vision-Instruct-mxfp8 \
  --quantize --q-bits 8 --q-group-size 32 --q-mode mxfp8 \
  --upload-repo mlx-community/North-Micro-Vision-Instruct-mxfp8

Usage

Cohere Compass support is available on the current MLX-VLM main branch. Install it directly from GitHub:

pip install -U "mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git"

Run vision-language inference:

mlx_vlm.generate \
  --model mlx-community/North-Micro-Vision-Instruct-mxfp8 \
  --image /path/to/image.jpg \
  --prompt "Describe this image." \
  --max-tokens 512 \
  --temperature 0.0

You can also pass an image URL to --image.

Notes

• MLX is optimized for Apple silicon.
• This repository changes the storage precision/quantization, not the source model architecture or intended behavior.
• Refer to the original model card for capabilities, limitations, licensing context, and responsible-use guidance.