SmolVLM2

SmolVLM2 is Hugging Face's second-generation tiny vision-language family, built on the Idefics3 architecture with a SigLIP image encoder and SmolLM2 as the text decoder. It ships in three sizes, 256M, 500M and 2.2B, with video-capable instruct variants, and trains on 3.3 million samples drawn from ten datasets spanning images, multi-image sets, video and text. The 256M video model needs only about 1.38 GB of GPU RAM for video inference, which is the whole point: multimodal understanding that fits on a laptop or a phone-class device.

The provider's evaluation has the 2.2B model at 52.1 on Video-MME, 55.2 on MLVU and 46.3 on MVBench, with the smaller sizes trailing by roughly ten points, plus solid document and chart scores (DocVQA 79.7, ChartQA 71.6 at 2.2B). Checkpoints are Apache 2.0 with Transformers, MLX and ONNX support, so deployment options are wide. As with any tiny VLM, expect fluent descriptions and decent OCR rather than the reasoning depth of models ten times its size.

• 2.2B (provider-reported): Video-MME 52.1, MLVU 55.2, MVBench 46.27; MathVista 43.9, MMMU 38.3, OCRBench 65.5, ChartQA 71.6, DocVQA 79.7
• 500M: Video-MME 42.2, MLVU 47.3, MVBench 39.73; 256M: Video-MME 33.7, MLVU 40.6, MVBench 32.7 (provider-reported)