The text recognition capability of the quantitative model is significantly inferior to that of the official model.
#1
by
xldistance
- opened
The official Qwen3-VL-30B-A3B model can perform OCR on text normally. The quantized model you created can only recognize partial text, and the responses it generates are modified versions of the input. Its ability to follow instructions has deteriorated.
The official Qwen3-VL-30B-A3B model can perform OCR on text normally. The quantized model you created can only recognize partial text, and the responses it generates are modified versions of the input. Its ability to follow instructions has deteriorated.
try this PR https://github.com/ggml-org/llama.cpp/pull/15474
xldistance
changed discussion status to
closed