OvisOCR3 — MLX (4bit)
Affine-quantized 4-bit (group size 64) MLX conversion of
ATH-MaaS/OvisOCR3 — a unified 0.8B
OCR / document-understanding model (document Markdown/JSON/layout parsing,
plain OCR, KIE, figure captioning, cross-page tables), fine-tuned from
Qwen/Qwen3.5-0.8B (qwen3_5
architecture).
Converted on 2026-11-14 with mlx-vlm 0.7.12:
python -m mlx_vlm convert \
--hf-path ATH-MaaS/OvisOCR3 \
--mlx-path OvisOCR3-4bit \
-q --q-bits 4 --q-group-size 64
Use
Requires mlx-vlm >= 0.3.12 (the first version that knows qwen3_5).
mlx_vlm.server --model fwgpiyawudk/OvisOCR3-mlx-4bit --port 8080 \
--min-pixels 200704 --max-pixels 8294400 # official 448*448 / 2880*2880
from mlx_vlm import load, generate
model, processor = load("fwgpiyawudk/OvisOCR3-mlx-4bit")
prompt = "<an official task prompt, e.g. document_markdown from prompts/>"
output = generate(model, processor, prompt=prompt, image="page.png",
max_tokens=6000, temperature=0)
- Prompts: the official task prompts ship in this repo's
prompts/folder (document_markdown,document_structured,document_layout,plain_ocr, ...) and define the output contract — use them verbatim. - Request shape: image first, then the prompt text; thinking disabled
(
enable_thinking=False— the source card's apply_chat_template passes it). - Sampling: greedy, temperature 0.0 (the source card's SamplingParams);
max_tokens 16384officially, 6000 in our eval runs.
Quantization note
No official 4-bit numbers exist for OvisOCR3. On the official Typhoon OCR V2 cards (same Qwen3.5 family) the 4-bit versions lost ~10 points on Thai pages (2B: 70.5 vs 80.6 MDPBench TH) while 8-bit matched bf16 — prefer 8-bit for Thai documents.