MiniCPM-V-4.7-1B-mxfp8

Converted from openbmb/MiniCPM-V-4.7-1B at revision 984a89aafcd5160164f4f0d04b4be0f4eec03cb8 with MLX-VLM main (e928101a144a3f11a27c7ef8d71b6d889b2d250c, version 0.7.8). Support was added in PR #2499.

MXFP8 weight quantization (mxfp8, group size 32). This is mixed precision: vision modules, visual mergers, embeddings, output heads, and linear attention remain BF16; recurrent A_log tensors retain their source precision. The 35B-A3B expert routers use 8-bit affine quantization.

See the original model card for the architecture, intended uses, and limitations. This conversion retains the source model's apache-2.0 license.

Usage

Install MLX-VLM from main with MiniCPM-V 4.7 support:

pip install -U git+https://github.com/Blaizzy/mlx-vlm.git@main
mlx_vlm.generate --model nativ-community/MiniCPM-V-4.7-1B-mxfp8 --image path/to/image.jpg --prompt "Describe this image." --max-tokens 200 --temperature 0

The native NumPy/Pillow processor requires neither PyTorch nor remote code. It supports text, images, multiple images, and video. The default visual downsampling is 16x; pass --processor-kwargs '{"downsample_mode": "4x"}' for finer detail.

Reproduce conversion

mlx_vlm.convert --hf-path openbmb/MiniCPM-V-4.7-1B --revision 984a89aafcd5160164f4f0d04b4be0f4eec03cb8 --mlx-path MiniCPM-V-4.7-1B-mxfp8 --dtype bfloat16 -q --q-bits 8 --q-group-size 32 --q-mode mxfp8

For the exact converter version, install MLX-VLM at commit e928101a144a3f11a27c7ef8d71b6d889b2d250c. Conversion options, per-module quantization overrides, and weight SHA-256 hashes are recorded in conversion.json.

Validation

The saved checkpoint passed strict reloading, quantization configuration checks, greedy text generation, and cat-image generation in both 16x and 4x downsampling modes on an Apple M3 Ultra with MLX 0.32.3. See validation.json for the outputs and environment. These are smoke checks, not a quality benchmark. Floating-point 4-bit formats can differ in output quality from affine quantization.