Historical reference for revision
cf22d043200e0173ee8939b3b211ed5ac815a12e. Results and validation below apply to the artifact digests recorded in that release, rather than the replacement weights.
Optimized Paddle throughput
This timing covers 2,078 unique ParseBench pages on one H100, with zero failed pages, at 3.088853185201328 pages/s. Four native Paddle pipeline instances share one vLLM server. Detection, crop recognition, assembly, and markdown conversion are timed; startup, image preload, and 16 warmups are excluded.
The context remains 16,384 tokens and the output cap remains 8,192 tokens.
The 4,096-token variants are excluded. See serving_contract.json and
benchmark_metrics.json; predictions are compressed in predictions.jsonl.gz.
Model-card accuracy uses the unchanged official published PaddleOCR-VL 1.6 results: overall 67.43, tables 67.77, charts 54.24, content 82.71, formatting 54.64, and Visual Grounding 77.80. These published accuracy scores are separate from the local timing measurement.