AWQ 4-bit (W4A16) quantization of baidu/Unlimited-OCR,
a 3B vision-language OCR model that pushes DeepSeek-OCR one step further (one-shot,
long-horizon document parsing). This repo quantizes the DeepSeek-V2 MoE text decoder with
activation-aware scaling (AWQ) while keeping the vision tower in BF16, so it stays a drop-in
transformers model.
⚠️ Runtime requirements. This is custom remote code, so load with
trust_remote_code=True, transformers 4.57.x, and compressed-tensors installed.
W4A16 (int4) runs on any CUDA GPU; compressed-tensors handles the 4-bit unpacking at load.
This quant
Scheme
W4A16 · int4 symmetric · group 128 · pack-quantized
vision tower (sam_model, vision_model), projector, token embeddings, lm_head, the MoE router gate, all norms, and the single dense layer-0 down_proj (width 6848 not divisible by group 128)
Unlimited-OCR uses the DeepSeek-OCR prompt vocabulary. The prompt must contain <image>;
prefix it with <|grounding|> whenever you also want bounding boxes for what was read.
Task
Prompt
Document → Markdown (layout-aware, with boxes)
`
\n<
Plain text OCR (just the text, no layout)
<image>\nFree OCR.
OCR with bounding boxes
`
\n<
Native Unlimited-OCR parse
<image>document parsing.
Parse a figure / chart / diagram
<image>\nParse the figure.
Describe the image (general VQA)
<image>\nDescribe this image in detail.
Find specific text (referring grounding)
`
\n<
Multi-page / PDF
<image>Multi page parsing. via model.infer_multi(...)
Resolution modes
base — base_size=1024, image_size=1024, crop_mode=False. Good default for normal pages.
gundam — base_size=1024, image_size=640, crop_mode=True. Tiles the page; use for dense
or large/high-resolution documents.
Understanding the output (grounding tokens)
With <|grounding|>, the model interleaves the recognized text with detection boxes:
Each [x1, y1, x2, y2] is the bounding box (top-left → bottom-right) of that span, in the
coordinate space of the model's input image. Drop the <|det|>...<|/det|> tags if you only want
text, or parse them to overlay boxes / rebuild layout. Without <|grounding|> you get plain text
(or Markdown) with no box tags.
Serving
The original model ships an SGLang wheel and a vLLM path (see the
base model card). W4A16 / compressed-tensors
weights load directly in runtimes with compressed-tensors support (e.g. vLLM); otherwise use the
transformers snippet above.
Task: multilingual OCR / document parsing — single image, multi-page, and PDF
(one-shot long-horizon parsing).
License: MIT (inherited from the base model).
How this was made
Unlimited-OCR is custom remote code whose forward only runs the vision tower when images are
passed, so AWQ calibration feeds text only (images=None), exercising the pure DeepSeek-V2
decoder. Per-layer AWQ mappings were built from the live module tree (attention
input_layernorm→q,k,v and v→o; MoE post_attention_layernorm→ every expert + shared-expert
gate/up, plus per-expert up→down). The fx-based "sequential" pipeline can't trace this custom
model, so the basic pipeline (real end-to-end forward + activation hooks) was used.
Verified
Loaded in transformers and run on a test document — OCR output matches BF16, e.g.:
4-bit weights trade a little accuracy for size; for the highest fidelity use the original BF16
model. For OCR, this AWQ build is effectively lossless on tested documents.
The vision encoder and MoE router stay BF16 (small, accuracy-sensitive).
English-/multilingual-text centric; verify critical fields on hard scans.
