PaddleOCR-VL-1.6-4bit

mlx-community/PaddleOCR-VL-1.6-4bit

4-bit MLX quantization of PaddlePaddle/PaddleOCR-VL-1.6, a 958M-parameter OCR / document-parsing VLM. Converted with mlx-vlm 0.6.10 (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/PaddleOCR-VL-1.6-4bit \
    --image document.png \
    --prompt "Transcribe this document to markdown." \
    --max-tokens 512

Quantization

Unlike most mlx-vlm conversions, the vision tower is quantized here: visual.* carries 84 .scales entries (27 qkv + 27 out_proj + 27 fc1 + position_embedding + 2 projector linears). The effective figure still exceeds 4 because the language model also keeps its embeddings, layernorms and the group-64 scales/biases in bf16.

Measured fidelity

Weight-level metrics compare dequantized weights against the bf16 source tensor by tensor, streamed from disk. No prompts, no sampling, hardware-independent. 27 fused visual.layers.N.self_attn.qkv tensors (source keeps q/k/v separate) are excluded -- they cannot be compared as single tensors; all other 185 quantized tensors are.

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. Six documents were rendered locally with exactly known content (invoice, clinical lab report, shipping label, receipt, motor-controller datasheet, and a second receipt rendered rotated 8 degrees to simulate a skewed photo scan), covering tabular line items, currency, dates, identifiers and an image-quality stressor. 76 required fields across the six.

• 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.

The 4-bit-vs-5-bit ordering is template drift, not a fidelity ranking

At 4 bits the variant transcribes all six documents perfectly; at 5 bits it confabulates tables on two of them (Item | Qty | Unit | Amount with duplicated amounts on the invoice, a hallucinated Name | Name / **Individuals** | **102** header on the lab report). The bf16 source itself is not exempt: it drops 24.00 and 16.00 on the invoice and concatenates PRIORITYOVERNIGHT.

This is template selection, verified, not assumed. Four controlled checks were run before publishing:

1. The bf16 control is byte-identical to the unquantized HF source on all six documents (loaded directly from the Hub, same mlx-vlm code path, greedy). The control is faithful — its field score is the base model.
2. All three variants agree on the first token's input (pixel values byte-identical) and diverge at the first generated token: the greedy path commits to a document template (markdown table vs plain lines) immediately, and a template mismatch costs whole fields, not characters.
3. The base model can produce the "perfect" transcript when steered: the bf16 control with the prompt "plain text, line by line, no tables" scores 0.987 (it was 0.961 before) — the content was always there, the default template was hiding it.
4. The ranking is not stable across prompts, which is the whole point: a real quality difference would survive a prompt change, and this one does not.

Three prompts, three different orderings: the 4-bit wins only under the default prompt, the 5-bit wins only under "markdown table", and under the steering prompt they tie within one field (PRIORITYOVERNIGHT against PRIORITY OVERNIGHT, a single dropped space in one document).

So the OCR ordering is not a quality ranking: greedy decoding is a discontinuous function of the weights, and quantization noise flips the template commitment at token 0 in a direction that is uncorrelated with bit width. The weight-level table above is the reliable ordering -- 5-bit reconstructs the source strictly better (24.2 dB vs 18.9 dB, the expected ~6 dB per bit). Six synthetic documents are too small a sample to rank variants on the OCR layer; treat the per-doc detail below rather than the aggregate as the signal.

Per-document field accuracy (missed fields in parentheses):

The rotated receipt (8 degrees, simulated photo) does not break any variant at the default prompt — 4-bit keeps all 14 fields there too.

What was not measured

• No distributional metrics (perplexity, top-1 agreement, KL). This is an OCR specialist; on plain chat text the bf16 source itself degenerates (asked for the capital of Japan it starts an unrelated derivation, and arithmetic prompts produce incoherent output), so an aggregate over ordinary prose would measure domain mismatch in the source, not quantization error.
• No standard OCR benchmark. No OmniDocBench, no DocVQA, no CER against a public dataset. The six documents are synthetic and rendered with a default font: exact ground truth, but not photographs, not handwriting, and not non-Latin scripts. The 8-degree rotation covers skewed scans in one controlled case, not the full variety.
• No judged quality, no layout-structure scoring (reading order, cell spanning), no throughput figures for any chip other than an M2 Pro.
• The 27 fused QKV tensors of the vision tower are not in the weight-level comparison (no single source tensor to compare them against).

Sibling variants

Credits

All credit for the underlying model belongs to PaddlePaddle. 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.