mmcheck 0.2.2

Check if a model supports multimodal inputs and what modalities it handles

mmcheck

Check if a model supports multimodal inputs.

pip install mmcheck

Quick start

from mmcheck import check

info = check("google/gemma-4-31B-it")
info.multimodal        # True
info.input_modalities  # ['text', 'image', 'video']
info.supports("image") # True
info.supports("audio") # False

CLI

mmcheck google/gemma-4-31B-it
# Model:      google/gemma-4-31B-it
# Multimodal: YES
# Inputs:     text, image, video
# Outputs:    text

mmcheck meta-llama/Llama-3-8B
# Multimodal: NO

mmcheck --json google/gemma-4-31B-it
mmcheck --offline gemma-4-31B-it

How it works

Three layers, checked in order:

1. Built-in registry — 30+ popular models (GPT-4o, Claude, Gemini, Llama, Qwen). Instant, no network.
2. HuggingFace Hub — fetches config.json, looks for vision_config, audio_encoder, architecture class names.
3. vLLM cross-reference — tags models with vLLM multimodal support status.

Modality	Detection
Image	vision_config, vision_tower, known VLM architectures
Audio	audio_config, audio_encoder, Whisper, Ultravox
Video	video_config, LLaVA-Next-Video, MiniCPM-V

Gated models

For gated HuggingFace models (401/403), mmcheck falls back to the public API metadata (tags, pipeline_tag). If you want full config inspection:

export HF_TOKEN=hf_...
mmcheck google/gemma-4-31B-it

Or in Python:

info = check("google/gemma-4-27b-it", token="hf_...")

License

MIT