base_model:
- xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
license: apache-2.0
tags: - comfyui
- qwen3.5
- w4a8
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic — ComfyUI W4A8
xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic as a single ComfyUI-native file, in the same format as Comfy-Org's qwen3.5_9b_w4a8_gptq.safetensors. Drop-in replacement for it in Load CLIP → Generate Text (prompt enhancers).
- Language layers: W4A8 (int4 + fp8 group-16 scales, convrot 256);
lm_headand embeddings: int8 convrot; vision tower and MTP head: bf16. - MTP speculative decoding works (
mtp: auto). - RTX 5090: ~280 tok/s, same as the stock W4A8 file.
Put it in ComfyUI/models/text_encoders/.