Home / Products / Image Generation / Vision

apple/ml-fastvlm

Apple's on-device vision-language model — 85× faster first-token output

Key features

• FastViTHD hybrid vision encoder outputting fewer tokens and faster encoding for high-res
• 0.5B variant: 85x faster time-to-first-token than LLaVA-OneVision-0.5B
• Larger variants with Qwen2-7B LLM outperform Cambrian-1-8B with 7.9x faster TTFT
• Inference supported on Apple Silicon (exported quantized models) and via demo iOS app
• Model zoo with 0.5B, 1.5B, 7B PyTorch checkpoints

Good for

• On-device vision language models on iPhone, iPad, Mac
• Efficient high-resolution image understanding with low latency

Stack & tags

Quick start

conda create -n fastvlm python=3.10
conda activate fastvlm
pip install -e .

From the project README on GitHub.

Similar open-source projects