Apple vision-language models

FastVLM

Ask questions about images. Start in your browser, then build with Python or WebGPU.

FastVLM combines Apple’s FastViTHD vision encoder with a language model to understand images and text together. Explore the 0.5B, 1.5B and 7B variants, choose the right checkpoint and try your own images.

Try FastVLM Models & downloads

Browser demo: FastVLM-0.5B · WebGPU + shader-f16 · about 1.1 GB on first run

Image understanding, on your device

Concept illustration of a laptop examining a photo of a tangerine and cup, with an image card and blue question bubble

Describe a scene, read text or ask about the details.

Open the playground

Choose a starting point

Model size, checkpoint format and runtime are separate choices. This playground runs a community ONNX conversion of 0.5B. Larger variants need a compatible local environment.

FastVLM 0.5B

Browser & small-device experiments

Start with image descriptions, text extraction or chart questions. Try a representative image before integrating it into a workflow.

Try the 0.5B demo

FastVLM 1.5B

Local development

Explore the middle-size model and its Stage3 checkpoint. Match the files to PyTorch, Transformers or Apple’s native app.

Read the Stage3 guide

FastVLM 7B

Higher published benchmark scores

Compare quality gains with memory and runtime requirements. Evaluate performance on the hardware you intend to use.

Compare models and downloads

From your first answer to your own app

Three differently sized compute modules beside the same photo, illustrating model selection

Models & downloads

Choose a model and download

Compare 0.5B, 1.5B and 7B, official benchmark scores and checkpoint formats.

Three image plates with the third highlighted in blue, illustrating checkpoint selection

Stage3 setup

Understand 1.5B Stage3

Choose the right training stage and files for PyTorch or Apple's native app.

A local development workspace with a laptop and an image of a tangerine and cup

Python & Transformers

Run FastVLM with Python

Set up Transformers, ask a question about an image and fix common first-run errors.

A browser image window above a graphics processor, illustrating local WebGPU inference

WebGPU tutorial

Build a browser image app

Download a complete starter with streaming, cancellation and JSON export.

Read benchmark claims in context

Published scores measure specific datasets and settings. They do not predict the quality or latency of every browser conversion. Compare sizes using the same benchmark row, then check your own images and requirements.

See scores and comparison conditions

What to check in the demo

Does the answer match the image, including numbers and labels?
Does it follow the requested output format?
How do the first load and later questions differ in latency?
Does it recover when a run is cancelled?

Questions before you start

Does my image leave this browser?

The image playground processes images and questions locally. It downloads model files from Hugging Face. Usage events contain action categories and timings rather than your content. The optional camera demo is a separate Hugging Face embed opened on request.

Why can’t I run it on this device?

The browser model needs WebGPU, shader-f16 and enough free GPU memory. A successful compatibility check does not guarantee enough memory. The Python guide provides another runtime.

Is Stage3 another model size?

Stage3 identifies a training checkpoint. The 1.5B label describes the model’s size. Choose both the checkpoint and the files for your runtime.

Is this an official Apple website?

AI IDE List is an independent developer resource and is not affiliated with Apple. FastVLM is developed by Apple researchers. Use the official sources below for releases, licenses and research details.

Official sources

Apple research Official repository & model zoo Apple model card Paper

Explore other Apple models and tasks on FastVLM.net: Apple model overview, browser tools and on-device VLM comparison.