HuggingFace unveils 2B compact VLM 'SmolVLM'

SmolVLM, developed by HuggingFaceTB, is a compact vision language model (VLM) family with 2B parameters. It is designed for deployment in local environments, browsers, and edge devices, offering excellent memory efficiency and inference speed.

The model is provided in three versions depending on use case:

• SmolVLM-Base: A base model for downstream fine-tuning.
• SmolVLM-Synthetic: A model fine-tuned on synthetic data.
• SmolVLM-Instruct: An instruction-following model ready to use for conversational applications.

All model checkpoints, datasets, training recipes, and tools are released under the Apache 2.0 license, allowing commercial use. It was trained using open-source datasets such as The Cauldron and Docmatix, and is integrated into the Transformers library for immediate use.