Regarding Q8_0 mmproj Quantization Issues in the Qwen3.5 Architecture
The Q8_0 quantization falls back to FP16 for the ffn_down weights because their dimension is not divisible by 32. You can resolve this issue by padding the dimension from 4,304 to 4,320.
You've been consistently releasing Q8_0 mmproj models, so I assume the FP16 fallback for ffn_up wasn't intentional—after all, a direct conversion would inevitably trigger this regression. The quality of a fully quantized Q8_0 version shouldn't differ noticeably from the BF16 version,and Q8_0 still has 128 times the dynamic range of FP16, haha.
https://huggingface.co/mradermacher/CoPaw-Flash-9B-GGUF This is an example. For the Qwen3.5 9B architecture, the full Q8_0 quantization should be 499 MB in size, yours is 624 MB because it includes fp16 ffn_down weights.