SmolVLM2 (HuggingFace): 256M to 2.2B VLMs for video understanding on consumer hardware
In one sentence HuggingFace releases SmolVLM2 in three sizes (2.2B, 500M, 256M) focused on video understanding. The 2.2B outperforms existing 2B models on Video-MME and needs about 5.2 GB of GPU memory for video inference; the 500M and 256M are presented as the smallest video language models ever released.
Until recently, getting an AI model to analyze video required large models and dedicated hardware. SmolVLM2, released by HuggingFace on February 20, 2025, brings video understanding to very small models.
There are three sizes: 2.2 billion, 500 million and 256 million parameters. According to HuggingFace, the largest one outperforms all existing 2B models on the Video-MME video benchmark; the 500M model has video capabilities very close to the 2.2B with less than a quarter of the parameters. The two small sizes are presented as the smallest video language models ever released.
The 2.2B needs about 5.2 GB of GPU memory for video inference and also runs on a free Colab. HuggingFace showed three demos: an iPhone app analyzing video locally with the 500M model, an experimental VLC player integration to search inside videos, and a highlight generator for long videos.
Companies
HuggingFace
Tools
SmolVLM2-2.2B, SmolVLM2-500M, SmolVLM2-256M
Tags
Sources