FASTVLM MODEL

FastVLM-1.5B

Official checkpoint overview, published benchmark scores and practical starting points.

BalancedBalanced

A middle-sized official variant that improves every score in Apple’s published table over 0.5B while remaining substantially smaller than 7B.

Best for: Balanced local inference and Apple Silicon development
Supported paths: Python · Apple Silicon · Workstation
Checkpoint: apple/FastVLM-1.5B

Download from Hugging Face

Download command

python -m pip install huggingface_hub
hf download apple/FastVLM-1.5B

What is FastVLM Stage3 1.5B?

FastVLM Stage3 1.5B is the 1.5B variant after the additional instruction-tuning stage described in Apple’s paper. Stage3 identifies the training checkpoint; 1.5B identifies the model size. It is not a separate installation package.

Stage2: visual instruction tuning. Apple releases a checkpoint at this point.
Stage3: further fine-tuning of Stage2 on high-quality instruction data that includes reasoning examples. Apple also releases this checkpoint.

Choose the files for your runtime

Runtime	Model files	Next step
Original PyTorch repository	Stage3 PyTorch ZIPDownload checkpoint ↗	Run predict.py after extracting the archive.
Official Apple Silicon app	Stage3 Apple Silicon ZIP · LLM int8Download checkpoint ↗	Use the app download script to place files in the correct folder.
Hugging Face Transformers	Use the compatible checkpoint in the Transformers guide. A Hugging Face repository and an Apple ZIP are not interchangeable.	Follow the complete image-to-answer example.

Run Stage3 1.5B on Mac or iPhone

Apple’s app README specifies macOS 15.2+ or iOS 18.2+. Install Xcode and select a supported target. Run these commands from the repository root; keep the model folder empty before the first download.

Apple Silicon · Stage3 1.5B

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.xcodeproj

Build and run in Xcode, select an image or camera input, and ask a short question. A successful run produces an answer and the app’s TTFT measurement. Exact text varies by image and prompt.

How much memory does it need?

There is no single verified minimum for all runtimes. Precision, image resolution, model buffers and generation length all affect peak memory. The int8 label describes the exported language-model weights, not total app memory. Start with 0.5B if 1.5B cannot load, and compare warm runs using the same image and prompt.

If the first run fails

Model not found: confirm the download completed into app/FastVLM/model before building.
Out of memory: close other GPU workloads, shorten the output, or switch to 0.5B.
Wrong format: use the Apple Silicon export for the app and the original PyTorch checkpoint for predict.py.

Complete Python / Transformers guide →

Sources checked September 12, 2026 · Paper · Model Zoo · App README

Published benchmark scores

Scores copied from Apple’s FastVLM model cards. Higher is better; compare only within the same benchmark row.

Benchmark	FastVLM-1.5B
AI2D	77.4
ScienceQA	94.4
MMMU	37.8
VQAv2	79.1
ChartQA	80.1
TextVQA	70.4
InfoVQA	59.7
DocVQA	88.3
OCRBench	70.2
RealWorldQA	61.2
SeedBench-Img	74.2

Recommended next steps

Read the Apple AMLR model license and model card.
Estimate hardware needs from the exact files and chosen precision.
Start with a representative image and prompt set.
Measure first-load, warm latency, memory and output quality on your target device.

Open implementation guides