FastVLM
AppleFastVLM is Apple's family of vision-language models built to answer quickly on high-resolution images. Its FastViTHD hybrid vision encoder emits fewer visual tokens and cuts encoding time, which shortens the wait for the first token of an answer. The models come in 0.5B, 1.5B and 7B sizes, the largest on Qwen2-7B, with exports for Apple Silicon and a demo iOS app that runs them on device.
Openness
2 high confidence- weights
- open (0.5B, 1.5B and 7B checkpoints and quantized Apple Silicon exports on the Hub, ungated)
- data
- closed (no training data released)
- code
- partial (inference, Apple Silicon export and the iOS demo app; training uses the LLaVA codebase)
- license
- Apple-AMLR (Apple Machine Learning Research Model license on the checkpoints; non-commercial research only. The repository code is under Apple's separate sample-code license)
The FastVLM checkpoints carry Apple's research-model license, which limits use to non-commercial scientific research and excludes any commercial product or service. Apple has released no training data, and training runs on the separate LLaVA codebase.
- https://huggingface.co/api/models?author=apple&search=fastvlm&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt recorded 2026-11-12
Six apple FastVLM repositories (0.5B, 1.5B, 7B and fp16, int8 and int4 exports), all gated false and tagged apple-amlr.
- https://raw.githubusercontent.com/apple-aiml-research/ml-fastvlm/main/LICENSE recorded 2026-11-12
Repository LICENSE: Apple's sample-code license for the Apple Software, copyright 2025 Apple Inc.
- https://raw.githubusercontent.com/apple-aiml-research/ml-fastvlm/main/LICENSE_MODEL recorded 2026-11-12
Apple Machine Learning Research Model license: use is granted "exclusively for Research Purposes", which "does not include any commercial exploitation, product development or use in any commercial product or service".
- https://raw.githubusercontent.com/apple-aiml-research/ml-fastvlm/main/README.md recorded 2026-11-12
README: "We use LLaVA codebase to train FastVLM variants. In order to train or finetune your own variants, please follow instructions provided in LLaVA codebase. We provide instructions for running inference with our models."
Adoption
2 high confidenceHugging Face downloads over the trailing 30 days for the three declared sizes, most of them for the 0.5B model. The quantized exports add comparatively little.
- https://huggingface.co/api/models?author=apple&search=fastvlm&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt recorded 2026-11-12
FastVLM-0.5B 16,089, FastVLM-1.5B 1,681 and FastVLM-7B 992 downloads in the trailing 30 days (18,762 together); the six repositories sum to 19,157.
Capability
2 high confidenceFastVLM reads documents and charts in a single image, level with Moondream. Neither its card nor its repository documents video or several images at once.
- https://huggingface.co/apple/FastVLM-7B/raw/main/README.md recorded 2026-11-12
FastVLM-7B card evaluations: DocVQA 93.2, ChartQA 85.0, InfoVQA 75.8, OCRBench 73.1 for FastVLM-7B; usage runs predict.py with one --image-file and a prompt.
Verified 2026-11-12