Falcon-OCR (MLX, bf16)
MLX-converted weights of
tiiuae/Falcon-OCR for inference
on Apple Silicon via mlx-vlm.
Source
- Base model: tiiuae/Falcon-OCR — Falcon Perception Team, Technology Innovation Institute (TII).
- License: Apache 2.0 (matches the upstream base model).
- Architecture: early-fusion vision-language model, 300M parameters.
Conversion details
- Tool:
mlx_vlm.convert(mlx-vlm 0.4.4). - Dtype:
bfloat16. - Source revision:
3a4d95a8b0008f7430df30a82cf35e6c3b6bcb66. trust_remote_code=True— the repository ships a customFalconOCRProcessor/FalconOCRForCausalLMthat is loaded via dynamic module import.
Known caveat
mlx_vlm.convert raises
AttributeError: 'FalconOCRProcessor' object has no attribute 'save_pretrained'
at the very end of the conversion step. The weights and tokenizer are written
successfully before the error — so the artifacts uploaded here are complete
and mlx_vlm.load(...) / docling's MlxVlmEngine can consume them.
Tracked upstream: https://github.com/Blaizzy/mlx-vlm/issues.
Usage
from mlx_vlm import load, generate
model, processor = load("mlx-community/Falcon-OCR-bf16", trust_remote_code=True)
output = generate(model, processor, prompt="", image=["path/to/page.png"])
print(output)
Attribution
All credit for the underlying model goes to the Falcon Perception Team at TII. Cite the model card for academic references.