SmolVLM: Hugging Face's 2B VLM built for a small memory footprint
In one sentence Hugging Face releases SmolVLM, a 2B-parameter vision-language model built on SmolLM2 1.7B that compresses images into few tokens and needs about 5 GB of GPU RAM. Checkpoints, datasets and recipes under Apache 2.0.
On November 26, 2024 Hugging Face publishes SmolVLM, a model that looks at images and answers questions about them, designed to use little memory. It has 2 billion parameters and comes in three versions: Base (a starting point for fine-tuning), Synthetic (fine-tuned on synthetic data) and Instruct (ready for end use).
Its strength is efficiency: an image with its question takes about 1,200 tokens, versus about 16,000 for Qwen2-VL, and the model runs with about 5 GB of GPU memory. Everything, including training data and recipes, is released under Apache 2.0, so it can also be used in commercial products.
Companies
HuggingFace
Tools
SmolVLM, SmolVLM-Base, SmolVLM-Synthetic, SmolVLM-Instruct
Tags
Sources