INT8 + ConvRot quantization of Kewk/Heretical-Qwen3.5-9B, for use as a ComfyUI text encoder.
| File | Link |
|---|---|
Heretic-Qwen3.5-9B-int8-convrot-Comfy.safetensors |
Download |
Accuracy
Weight error vs the bf16 source (relative L2, over all 200 quantized layers): 0.0090
about 2.9x lower than per-tensor FP8-e4m3 (0.026).