MiniCPM-V 4.6

MiniCPM-V 4.6 is an edge-oriented vision-language model for efficient image, multi-image, and video understanding. It combines a SigLIP2-400M vision encoder with a Qwen3.5-0.8B language model and supports mixed 4x/16x visual token compression so applications can trade finer detail for faster inference.

Model

Model	Repository
bf16	mlx-community/MiniCPM-V-4.6-bf16
4-bit	mlx-community/MiniCPM-V-4.6-4bit
5-bit	mlx-community/MiniCPM-V-4.6-5bit
8-bit	mlx-community/MiniCPM-V-4.6-8bit
NVFP4	mlx-community/MiniCPM-V-4.6-nvfp4
MXFP4	mlx-community/MiniCPM-V-4.6-mxfp4
MXFP8	mlx-community/MiniCPM-V-4.6-mxfp8

Details

Architecture	SigLIP2-400M vision encoder + Qwen3.5-0.8B LLM
Visual Compression	Mixed 4x/16x downsampling, with 16x as the default efficient mode
Modalities	Text, image, multi-image, video
Tasks	Image description, visual question answering, OCR, document understanding, visual reasoning, video understanding
License	Apache-2.0
Official Card	openbmb/MiniCPM-V-4.6

CLI Usage

Image inference:

python -m mlx_vlm.generate \
    --model mlx-community/MiniCPM-V-4.6-bf16 \
    --image path/to/image.jpg \
    --prompt "Describe this image" \
    --max-tokens 200 \
    --temperature 0.0

Video inference:

python -m mlx_vlm.generate \
    --model mlx-community/MiniCPM-V-4.6-bf16 \
    --video path/to/video.mp4 \
    --fps 1 \
    --prompt "Describe what happens in this video." \
    --max-tokens 300 \
    --temperature 0.0 \
    --processor-kwargs '{"max_num_frames": 32, "stack_frames": 1, "max_slice_nums": 1, "use_image_id": false}'

Use downsample_mode="4x" when a task needs finer visual detail:

python -m mlx_vlm.generate \
    --model mlx-community/MiniCPM-V-4.6-bf16 \
    --image path/to/image.jpg \
    --prompt "Read all visible text." \
    --max-tokens 300 \
    --temperature 0.0 \
    --processor-kwargs '{"downsample_mode": "4x", "max_slice_nums": 36}'

Python Usage

from mlx_vlm import generate, load
from mlx_vlm.prompt_utils import apply_chat_template

model_path = "mlx-community/MiniCPM-V-4.6-bf16"
model, processor = load(model_path)

images = ["path/to/image.jpg"]
prompt = "Describe this image"

formatted_prompt = apply_chat_template(
    processor,
    model.config,
    prompt,
    num_images=len(images),
)

result = generate(
    model,
    processor,
    formatted_prompt,
    image=images,
    max_tokens=200,
    temperature=0.0,
)
print(result.text)

Architecture

• Vision: SigLIP2-400M image encoder with MiniCPM-V patch packing and a window-attention merger inserted in the vision stack.
• Language: Qwen3.5-0.8B language model backbone for text generation and multimodal reasoning.
• Compression: The default 16x visual downsampling path is optimized for throughput and lower token cost. The 4x mode keeps more visual tokens for detail-heavy OCR, charts, documents, and small objects.
• Efficiency: The official model card reports more than 50% lower visual encoding FLOPs from the LLaVA-UHD v4 technique, plus roughly 1.5x token throughput compared with Qwen3.5-0.8B.
• Deployment: The upstream model is designed for broad edge deployment, including iOS, Android, and HarmonyOS.

Notes

• The official card highlights strong single-image, multi-image, and video understanding while keeping the model small enough for mobile-oriented use.
• For evaluation-style prompts, keep thinking disabled when the chat template exposes that option. The official Transformers examples use enable_thinking=False for this mode.
• downsample_mode controls placeholder counts during preprocessing and the vision encoder path during generation. Keep it consistent across processing and generation when using low-level APIs.
• For video, the official card recommends max_slice_nums=1 and use_image_id=False; the MLX video path follows the same practical defaults in the example above.
• The model weights and upstream code are released under Apache-2.0 by OpenBMB.