SmolVLM

Hugging Face built SmolVLM to prove vision language models can be small: the image models span 256 million, 500 million, and roughly 2 billion parameters, and SmolVLM2 carries the same three sizes over to video understanding. Each model pairs a SigLIP vision encoder with a SmolLM2 language backbone and uses aggressive pixel shuffle compression, encoding a test prompt plus one image in about 1.2k tokens where comparable models need 16k, so the 2B model fits in roughly 5 GB of GPU memory and the 256M variant, among the smallest capable VLMs released, runs inference on one image in under 1 GB. Everything is unusually transparent for the category: weights, the Cauldron and Docmatix training datasets, and the training recipes are published under Apache 2.0, so commercial use is unrestricted. The models load through the transformers library, run on CPU or GPU, and serve as bases for fine-tuning. Developers use them for document QA, captioning, and visual reasoning on phones and edge boxes, with code and resources kept alongside the SmolLM language models in the smollm repository.