SmolVLM2-2.2B

SmolVLM2-2.2B is the largest variant of SmolVLM2, the compact vision-language model family released February 20, 2025 by Hugging Face's TB Research team. Designed to make video understanding accessible on consumer hardware, the 2.2B model runs in about 5.2GB of GPU VRAM, fitting on an RTX 3060, a MacBook Pro M2, or even the free Google Colab T4 tier. On Video-MME, the standard long-form video understanding benchmark, it outperforms every existing model at the 2B scale according to Hugging Face's reporting.

The model compresses image tokens efficiently to process up to 50 frames per video, supports multi-image conversations through a chat template API, and ships with MLX-ready builds for Apple Silicon from day zero. The family also includes 500M and 256M sizes for tighter budgets. It is released under Apache 2.0 as HuggingFaceTB/SmolVLM2-2.2B-Instruct and has become a default choice for local video summarization pipelines, OCR-focused edge deployments, and on-device visual Q&A.