PaddleOCR-VL-1.5

Benchmark

Model	Framework	Scene	TTFT	Token/s	Sample	Notes
PaddleOCR-VL-1.5 (f16)	llama.cpp	OCR / ReadingGeneration	310 ms	94	-	Can recognize complex layouts, but may exhibit hallucinations when encountering large blocks of repeated text.
PaddleOCR-VL-1.5-GGUF-Q4_K_M.gguf	llama.cpp	ReadingGeneration	-	199.66	307 tokens / 1.5 s	More prone to hallucinations compared to f16
PaddleOCR-VL-1.5-GGUF-Q8_0.gguf	llama.cpp	ReadingGeneration	-	176.82	2012 tokens / 11 s	Balanced speed and stability; recommended for use with this precision.

Notes

• This model primarily provides pure OCR capabilities.
• To achieve layout analysis and Markdown output, it must be used in conjunction with other models.
• Container image: registry.lazycat.cloud/x/lzc-aipod-paddleocr-vl:2017cd2