mlx-community/Unlimited-OCR-2-bf16

baidu/Unlimited-OCR-2 converted to MLX with mlx-vlm: bfloat16 weights (no quantization). The SAM and CLIP vision towers and the per-layer DASH router stay in bfloat16.

Refer to the original model card for details on the model.

Use with mlx-vlm

Unlimited-OCR-2 support (page2d RoPE and DASH block-sparse attention) is being upstreamed to mlx-vlm. Until it is released, install from the branch that adds it:

pip install -U "git+https://github.com/shuuul/mlx-vlm.git"

# One page (gundam mode: 1024 global view + up to 16 local crops)
mlx_vlm.generate --model mlx-community/Unlimited-OCR-2-bf16 --image page.png \
    --prompt "Multi page parsing." --max-tokens 8192 --temperature 0

# Several pages in one request (one 1024 base view per page)
mlx_vlm.generate --model mlx-community/Unlimited-OCR-2-bf16 --image p1.png p2.png \
    --prompt "Multi page parsing." --max-tokens 16384 --temperature 0

The output uses <PAGE>, <BLOCK>label[x0, y0, x1, y1]<FORK>text and cross_* / append_* labels for blocks that continue across a page break; the upstream repository ships a converter to Markdown.