SmolVLM 2B Vision-Language Model enables edge deployments with low memory footprint

AI Impact Summary

SmolVLM releases a 2B Vision-Language Model with a strong emphasis on memory efficiency, enabling on-device inference and edge deployment. The model switches the language backbone to SmolLM2 1.7B and employs aggressive 9x visual-information compression via pixel shuffle (patches 384x384, inner 14x14 with a SigLIP backbone), achieving a minimum GPU RAM requirement around 5.02 GB. As an open-source Apache 2.0 project with ready-made variants (Base, Synthetic, Instruct) and open training pipelines, it lowers the barrier to local multimodal workloads and could shift some workloads away from cloud inference to client devices.