SmolVLM Grows Smaller – Introducing 256M & 500M Models

AI Impact Summary

SmolVLM is releasing significantly smaller models (256M and 500M parameters) alongside existing larger models, representing a shift towards efficiency and accessibility. This move leverages optimizations like a smaller vision encoder, data mixture adjustments, and tokenization improvements to achieve performance comparable to the 2B model while dramatically reducing memory requirements. This change enables deployment on constrained devices and consumer laptops, opening up new use cases for multimodal inference.