SmolVLM-256M

SmolVLM-256M is Hugging Face's tiny vision-language model, announced January 23, 2025, and at release billed as the smallest VLM in the world with just 256 million parameters. It shrinks the SmolVLM 2B design with a 93M-parameter SigLIP-based vision encoder and a 135M language model, cutting the vision tower to less than a quarter of its predecessor.

The model handles image captioning, visual question answering, document Q&A, and basic visual reasoning on consumer laptops with under 1GB of RAM, and supports WebGPU for in-browser inference. Four checkpoints shipped at launch: base and instruction-tuned variants at both 256M and 500M sizes, loadable in transformers, MLX, and ONNX.

Weights are open under the Apache 2.0 license. The 256M model trades caption richness for footprint compared to its larger siblings, but it has found a niche in edge vision: developers have ported it to Android GPUs, IP security cameras, and Jetson-class hardware.