SmolVLM2 Captioner

A CLAMS app that captions video TimeFrames with the SmolVLM2-2.2B-Instruct multimodal model.

This app ships only the 2.2B-Instruct variant: the largest and most general-purpose model in the SmolVLM2 family. The smaller (256M and 500M) SmolVLM2 releases are post-trained specifically for video-QA tasks and we do not expect them to generalize well, given their size.

The canonical user-facing documentation (app description, runtime parameters, I/O specs, GPU memory estimates) lives in metadata.py, rendered at the CLAMS App Directory. For general CLAMS app usage, see the CLAMS App Manual.

System requirements

• CUDA-capable NVIDIA GPU (see est_gpu_mem_min / est_gpu_mem_typ in metadata.py for VRAM expectations).
• ffmpeg (used by mmif-python[cv]).