Hugging Face just dropped a family of video-understanding models so small you can run them on a phone, and they’re not terrible. SmolVLM2 comes in three sizes — 2.2 billion, 500 million, and a frankly absurd 256 million parameters — with the two smaller variants claiming the title of smallest video-language models ever released. The company says the 500M model gets you video comprehension nearly on par with the 2.2B version at less than a quarter the size.
The headliner is the 2.2B model, which the team says now outperforms every existing 2B-class model on the Video-MME benchmark, a grueling test spanning 900 videos from 11-second clips to hour-long films with subtitles and audio. It’s also better than its predecessor at math-with-images, reading text in photos, and parsing complex diagrams. The team credits improved data mixture strategies from the Apollo paper for the balanced video-and-image performance.
But the real statement here is the 256M model. Nobody has shipped a video-capable model this small before. Hugging Face calls it an experimental release — a “what if” project — taking cues from IBM’s earlier work fine-tuning the base SmolVLM-256M. The implication is clear: they’re daring developers to try specialized fine-tuning runs that would be impractical with larger models.
All three models are available immediately with full MLX support, meaning Swift and Python developers can start building on Apple Silicon from day zero. The team also released three demo apps — an iPhone video understanding tool, a VLC media player integration, and a video highlight generator — suggesting they see this less as a research artifact and more as a developer toolkit. For anyone who’s watched the video AI space obsess over ever-larger foundation models, seeing competitive results from something that fits in a free Google Colab instance feels like a genuine shift in what’s possible on the edge.
