Shypd.aiThe largest open index of AI toolsml-fastvlm is an Open Source tool that provides an efficient vision encoding for Vision Language Models. It features FastViTHD, a hybrid vision encoder designed to reduce encoding time for high-resolution images.
Trending
About
ml-fastvlm is the official implementation of "FastVLM: Efficient Vision Encoding for Vision Language Models," a project presented at CVPR 2025. This tool introduces FastViTHD, a novel hybrid vision encoder that significantly reduces the number of tokens and encoding time for high-resolution images. Its smallest variant boasts 85x faster Time-to-First-Token (TTFT) and a 3.4x smaller vision encoder compared to LLaVA-OneVision-0.5B. Larger variants, utilizing the Qwen2-7B LLM, outperform recent models like Cambrian-1-8B with a 7.9x faster TTFT. The repository provides instructions for training, fine-tuning, and running inference, including support for Apple Silicon and Apple devices like iPhone, iPad, and Mac, with a demo iOS app available.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
