OvisOCR2-4bit

OvisOCR2-4bit

4-bit MLX quantization of ATH-MaaS/OvisOCR2, an 853M-parameter OCR / document-parsing VLM. Converted with mlx-vlm 0.6.8 (mlx 0.32.0) for inference on Apple Silicon.

Fidelity was measured against the unquantized bf16 source, not assumed. The numbers below and the method that produced them are in the tables that follow.

pip install mlx-vlm
python -m mlx_vlm generate \
    --model mlx-community/OvisOCR2-4bit \
    --image document.png \
    --prompt "Transcribe this document to markdown." \
    --max-tokens 512

Quantization

Effective bits per weight exceeds the nominal 4 because mlx-vlm quantizes only the language model and leaves the vision tower in bf16 by design. Verified, not assumed: of 153 vision tensors, zero carry a .scales entry.

Note this model ties its input and output embeddings, and the embedding is 33.8% of all quantized parameters, a much larger share than in a typical 7B. Weight error is therefore dominated by the embedding rather than the attention layers.

Measured fidelity

Weight-level metrics compare dequantized weights against the bf16 source tensor by tensor, streamed from disk. No prompts, no sampling, hardware-independent.

Measured on an M2 Pro / 32 GB, greedy decoding.

OCR accuracy vs ground truth

Because OCR has ground truth, transcription can be scored directly rather than inferred from token distributions. Five documents were rendered locally with exactly known content (invoice, clinical lab report, shipping label, receipt, motor-controller datasheet), covering tabular line items, currency, dates and identifiers.

• Field accuracy: required strings present, exact match (format-sensitive).
• Content accuracy: the same, ignoring markup and whitespace.
• Numeric recall: ground-truth numbers present, ignoring markup. For an invoice or a datasheet the numbers are the payload.

Every variant preserves 100% of content and 100% of the 77 ground-truth numbers. At 6 and 8 bits the transcriptions are character-identical to bf16 (CER 0.0000 across all five documents), not merely close.

Why 4-bit shows CER 0.331 but loses no content

The 4-bit variant scores 0.968 on strict field accuracy and 0.331 CER, which looks like degradation. It is not. At 4 bits the model emits an HTML table where bf16 emits markdown:

bf16 : Widget A 2 12.00 24.00
4-bit: <tr><td>WidgetA</td><td>2</td><td>12.00</td><td>24.00</td></tr>

Every figure is correct. The only strict-match failures in all five documents are WidgetA against Widget A, a single dropped space. Content accuracy and numeric recall are both 1.000.

CER against bf16 treats the bf16 output as ground truth, so it penalises a valid alternative rendering the same way BLEU penalises paraphrase. Reported here for continuity, but it measures drift, not quality. If you post-process into structured data, 4-bit is usable and is the fastest of the three; if you need byte-stable markdown, use 6-bit.

What was not measured

• No distributional metrics (perplexity, top-1 agreement, KL). They were run and are deliberately not reported. This model is an OCR specialist, and on general chat text the bf16 source itself degenerates (asked for the capital of Japan it emits Tokyo / Nara / Nara / Nara ...). Its perplexity spans 49.3x across passages of ordinary prose, so an aggregate over them measures domain mismatch in the source, not quantization error. Per-passage ratios do improve monotonically from 4 to 8 bits.
• No standard OCR benchmark. No OmniDocBench, no DocVQA, no CER against a public dataset. The five documents are synthetic and rendered with a default font: exact ground truth, but not photographs, not handwriting, not skewed scans, and not non-Latin scripts.
• No judged quality, no layout-structure scoring (reading order, cell spanning), no throughput figures for any chip other than an M2 Pro.

Sibling variants

Credits

All credit for the underlying model belongs to ATH-MaaS. This is a format conversion and quantization; no training or fine-tuning was performed. See the source model card for license, intended use and limitations.