Skip to content
FastVLM.net

FASTVLM GUIDE · Models

FastVLM on Hugging Face

Find the official Apple model collection, choose 0.5B, 1.5B or 7B, and review each model card before downloading.

All FastVLM guides

Before you run

Use Python 3.10+ in a new environment. This example pins Transformers 5.0.0, which includes native FastVLM support. Install a PyTorch build appropriate for your hardware. The script chooses CUDA, Apple MPS or CPU; CPU inference can be slow.

Match the checkpoint to the library

apple/FastVLM-* contains Apple’s released model files. The native Transformers example uses KamilaMila/FastVLM-0.5B, the community conversion referenced by Hugging Face documentation. The browser uses onnx-community/FastVLM-0.5B-ONNX. Choose the runtime first, then its compatible files.

Troubleshooting

  • ImportError for FastVlmForConditionalGeneration: check that the command uses the environment where Transformers 5.0.0 was installed.
  • Model download fails: check Hugging Face connectivity and disk space; the first load downloads weights.
  • Out of memory: use 0.5B, close other GPU workloads and reduce output length.
  • Prompt copied into the answer: decode only tokens generated after input_ids, as this example does.
Stage2 / Stage3 and Apple Silicon files →

Source checked September 12, 2026. Runtime performance depends on your device.

Step by step

  1. 1

    Open Apple’s FastVLM collection.

  2. 2

    Choose a model based on footprint and published benchmark needs.

  3. 3

    Read the Apple AMLR model license and model card.

  4. 4

    Download with the Hugging Face CLI or your chosen library.

Example

Models 
python -m pip install huggingface_hub
hf download apple/FastVLM-0.5B --local-dir ./models/apple-fastvlm-0.5b
# For native Transformers inference, use the compatible checkpoint
# and complete script in /docs/fastvlm-transformers.

Verify against the primary source

APIs, model files and dependency versions can change. Treat this page as an implementation map, then confirm the current command and license in the linked official documentation.

Open official source