Variant mixed. See the project's bench/RESULTS.md for the precision benchmark. The mixed variant keeps the vision tower, embeddings, lm_head and the 18 Gated-DeltaNet layers at bf16 and quantizes only the bulk linears — strictly more precise than the existing 8-bit port.

python

from mlx_vlm import load, generate
model, processor = load("aglaia-models/surya-ocr-2-mlx")

Changes from the original

Weights converted to MLX format and quantized (mixed). No fine-tuning.

License & attribution

Derivative of datalab-to/surya-ocr-2, under the AI Pubs OpenRAIL-M License (Modified). This derivative is released under the same license (share-alike); the use-based restrictions carry over. Credit: Datalab (https://github.com/datalab-to/surya).