8-bit MLX quantization of baidu/Unlimited-OCR -- a long-horizon document OCR / parsing model (DeepEncoder vision + DeepSeek-V2 MoE decoder) -- running natively on Apple Silicon via mlx-vlm.

At a glance


Source	`baidu/Unlimited-OCR` @ `ee63731b6461c8afcdcc7b15352e7d2ffecc2ead`
Architecture	DeepEncoder (SAM-ViT + CLIP-L) vision + DeepSeek-V2 MoE decoder (12 layers, 64 experts)
Format	MLX (Apple Silicon native), loads via `mlx-vlm`'s `deepseekocr`
Quantization	8-bit, group-size 64 (9.41 bits/weight effective)
Disk size	3.7 GB

Verification (2026-07-12, M-series Mac)

Converted with mlx-vlm 0.3.9 and document-OCR smoke-tested locally before publishing -- this pack correctly extracts all fields from a test invoice (number, date, bill-to, line item, amount, status) with bounding-box grounding.

Sibling variants

Variant	Size	Status
Unlimited-OCR-MLX-8bit	3.7 GB	public
Unlimited-OCR-MLX-6bit	3.0 GB	public

Usage

Requires mlx-vlm >= 0.3.9. This repo already carries the small config shim so mlx-vlm routes it through its deepseekocr implementation (upstream model_type: unlimited-ocr -> deepseekocr; processor -> DeepseekVLV2Processor):

python

from mlx_vlm import load, generate
model, processor = load("majentik/Unlimited-OCR-MLX-8bit")
out = generate(model, processor, "<image>document parsing.", ["page.png"], max_tokens=2048)
print(out)

Prompts follow the upstream convention: <image>document parsing. for single documents; the model emits text with <|det|>...<|/det|> bounding-box grounding.

Intended use

Document OCR, layout parsing, and text extraction (English + Chinese) locally on Apple Silicon. See the base model card for benchmarks and long-horizon / multi-page usage.

Conversion notes

Loadable via mlx-vlm's existing deepseekocr module -- Unlimited-OCR shares the DeepEncoder + DeepSeek-V2 architecture with DeepSeek-OCR. The only changes from upstream are two config fields (model_type, processor_class/sft_format) so the stock loader routes it; the weights are unmodified beyond quantization.

Reproduce

This pack is produced from upstream baidu/Unlimited-OCR with a small config shim so mlx-vlm routes it through its existing deepseekocr implementation, then converted with mlx_vlm.convert:

Patch config.json: model_type: unlimited-ocr -> deepseekocr.
Patch processor_config.json: processor_class -> DeepseekVLV2Processor, sft_format -> deepseek.
Convert:

bash

python -m mlx_vlm convert \
  --hf-path <patched dir> \
  --mlx-path <out> \
  -q --q-bits 8 --q-group-size 64

The weights themselves are unmodified beyond quantization; only the two config fields above are changed so the stock mlx-vlm loader recognizes the architecture (Unlimited-OCR shares the DeepEncoder + DeepSeek-V2 MoE decoder architecture with DeepSeek-OCR).

License

MIT, inherited from the upstream model.