Fastest
FastVLM-0.5B
Best entry point for WebGPU, browser demos and constrained edge devices.
Browser · Mobile · Edge
View modelPLAYGROUND · MODELS · DEPLOYMENT
Run FastVLM in Your Browser
Try real-time vision-language inference with WebGPU, compare FastVLM-0.5B, 1.5B and 7B, and follow practical deployment guides for browser, Python, Mac and iPhone.
- Runs in browser
- WebGPU acceleration
- FastVLM-0.5B
FastVLM-0.5B
See what FastVLM sees
prompt › Describe what is happening in front of the camera.
FastVLM FastVLM processes the visual stream locally in the browser and generates a live description.
FastViTHDTransformers.jsWebGPUMLXCore ML
Start here
A real FastVLM WebGPU playground
Open the embedded FastVLM demo, grant camera access when you are ready, and run the 0.5B model through Transformers.js and WebGPU. The original demo stays available on its own focused page.
apple/FastVLM-0.5BWebGPU
Choose a model
Three FastVLM sizes, one efficient vision encoder
Start with 0.5B for browser and edge experiments, move to 1.5B for a stronger balance, or use 7B when benchmark quality matters most.
Balanced
FastVLM-1.5B
A practical middle ground for local development and Apple Silicon.
Mac · Workstation
View model Best quality
FastVLM-7B
The strongest official variant across the published evaluation table.
Desktop · Native
View modelFastVLM explained
High-resolution vision without the usual token cost
FastVLM is an open vision-language model from Apple researchers. Its FastViTHD hybrid vision encoder reduces visual tokens and encoding latency, improving the accuracy–latency trade-off for high-resolution images and on-device use.
Apple-reported comparison In Apple’s reported comparison at the highest resolution, FastVLM-0.5B achieved up to 85× faster time-to-first-token than LLaVA-OneVision-0.5B while using a 3.4× smaller vision encoder.
Image
FastViTHD
Visual tokens
LLM
Response
Official results
FastVLM benchmark snapshot
Published evaluation scores from Apple’s Hugging Face model cards. Higher is better; benchmark protocols differ, so compare within each row.
| Benchmark | 0.5B | 1.5B | 7B |
|---|---|---|---|
| AI2D | 68.0 | 77.4 | 83.6 |
| ScienceQA | 85.2 | 94.4 | 96.7 |
| MMMU | 33.9 | 37.8 | 45.4 |
| ChartQA | 76.0 | 80.1 | 85.0 |
| TextVQA | 64.5 | 70.4 | 74.9 |
| DocVQA | 82.5 | 88.3 | 93.2 |
Source: Apple FastVLM model card Hugging Face
Deploy your way
Run FastVLM where your application lives
Browser
Use the 0.5B WebGPU demo without installing a local runtime.
Open playgroundPython
Load official Hugging Face checkpoints for local inference.
Python guideMac
Use Apple Silicon resources including MLX-compatible checkpoints.
Mac guideiPhone
Explore Apple’s official iOS/macOS demo and Core ML resources.
iPhone guidePopular guides
From first run to local deployment
Make the trade-off explicit
Compare FastVLM before you deploy
See how model sizes trade accuracy for footprint, and read a source-qualified comparison with LLaVA-OneVision.
FAQ
FastVLM questions, answered
Can FastVLM really run in a browser?
Yes. The linked FastVLM-0.5B demo uses Transformers.js and WebGPU to run inference directly in a compatible browser. Initial model download and compilation can take time.
Which FastVLM model should I start with?
Use 0.5B for the browser and constrained devices, 1.5B for a balanced local setup, and 7B when published benchmark quality is the priority.
Is FastVLM.net the official Apple website?
No. FastVLM.net is an independent developer resource. Apple’s research page, GitHub repository and Hugging Face organization are the official sources linked throughout this site.
Does every FastVLM setup run fully on-device?
No. FastVLM supports on-device deployment, but privacy and data flow depend on the runtime and application you choose. The WebGPU playground is designed for in-browser inference.
Primary sources