SmolVLM

SmolVLM is Hugging Face's family of tiny vision-language models (256M, 500M, and 2.2B parameters), built on the Idefics3 architecture with a SigLIP vision encoder, aggressive pixel-shuffle visual token compression, and SmolLM2 as the language backbone. Released January 2025 under Apache 2.0, the 256M variant runs on devices with under ~1GB of RAM and still beats the 80B-parameter Idefics on the AI2D science-diagram benchmark. SmolVLM2 adds multi-image and video input support.

The honest picture: tiny is tiny — on MMMU the 256M scores 28.3 and even the 2.2B only reaches 38.3, far below frontier VLMs. These are for on-device captioning, document QA, and cheap batch processing, not for hard visual reasoning. Context is limited (8K tokens on the 256M) and video support is basic. But for the specific job of "good enough vision on hardware that costs nothing," nothing else in the open world touches them.