On This Page6 sections
FastVLM 1.5B Stage3 combines a model size with a training checkpoint. “1.5B” names the size; “Stage3” identifies the later instruction-tuning checkpoint described in the research. Your next decision is the runtime: Apple's original PyTorch implementation, its native Apple Silicon app, or a compatible Transformers conversion.
See the model comparison and downloads for all three sizes, or try 0.5B online before setting up a local environment.
Stage3
Apple releases both Stage2 and Stage3 checkpoints. Stage2 follows visual instruction tuning. Stage3 continues training with additional high-quality instruction data, including reasoning examples. The stage identifies which weights you download; it does not change a ZIP archive into an application installer. See the paper and official checkpoint list.
Use the stage that matches the experiment or deployment instructions you intend to reproduce. Record it alongside model size, precision and runtime when comparing results.
Pick the correct download
| Where you will run it | Files to use | How they are consumed |
|---|---|---|
| Original PyTorch implementation | Stage3 1.5B PyTorch ZIP | Extract the archive and pass its model directory to the repository's prediction script. |
| Apple's native app on Mac/iPhone | Stage3 1.5B Apple Silicon ZIP | Use the official app download script to put the exported files in the expected location. |
| Hugging Face integration | Apple's 1.5B model card or a specifically supported conversion | Follow that checkpoint's usage code and compatible library version. |
The llm.int8 filename describes the exported language-model weights. It is not a measurement of the complete app's memory consumption. The official export instructions explain the Apple Silicon path.
Run the Apple app
Install Xcode and check the supported operating systems and build settings in the app README. The original guide specifies macOS 15.2+ or iOS 18.2+; check the current app requirements when selecting your target.
From a fresh checkout:
git clone https://github.com/apple-aiml-research/ml-fastvlm.git
cd ml-fastvlm
chmod +x app/get_pretrained_mlx_model.sh
app/get_pretrained_mlx_model.sh --model 1.5b --dest app/FastVLM/model
open app/FastVLM/FastVLM.xcodeprojStart with an empty model destination rather than mixing exports from several model sizes. In Xcode, select the supported device or Mac target, configure signing when required, and build the application. Choose an image or camera input and ask a short question. Review the answer against the image and inspect the app's timing display.
These commands follow the official app workflow. The app was not built on every supported device as part of this article; a successful download alone does not validate the Xcode build or model output.
Use the original PyTorch checkpoint
In the official repository's configured Python environment, point predict.py at the extracted checkpoint directory:
python predict.py \
--model-path /path/to/extracted-stage3-1.5b \
--image-file /path/to/image.png \
--prompt "Describe this image briefly."Use the repository setup instructions for its dependencies. Do not pass the app's Apple Silicon export to this command. For a complete native Transformers image-question example using the supported 0.5B conversion, follow Run FastVLM with Python.
Memory and first-run troubleshooting
| Symptom | What to check |
|---|---|
| Model files cannot be found | Confirm that the download completed in app/FastVLM/model and that the app target includes the expected files. |
| Checkpoint loads with missing or incompatible weights | Check the training stage, conversion and runtime together. Avoid mixing model directories. |
| GPU memory error | Close competing workloads, shorten generation or start with 0.5B. Image resizing does not shrink the model weights. |
| Xcode build failure | Compare the error with the app's current README, target OS, signing and dependency settings. |
| Fluent but incorrect answer | Compare it with the original image. Test several representative cases; successful execution does not establish correctness. |
There is no single verified minimum-memory figure covering all FastVLM 1.5B deployments. Precision, image resolution, temporary buffers and generation length affect peak use. Measure on the intended device and keep those conditions with your results.
Continue
- Compare all FastVLM models and official scores
- Run the complete Python example
- Build a WebGPU image app
- Return to the FastVLM hub
Article revised September 23, 2026. Checkpoint and app instructions are linked to their primary sources; no new 1.5B performance measurement is claimed here.
Continue Reading
More articles connected to the same themes, protocols, and tools.
Referenced Tools
Browse entries that are adjacent to the topics covered in this article.
