MisterNatural/Qwen3.8-9B-Distill-ComfyUI

Qwen3.8-9B-Distill — ComfyUI Models

ComfyUI-adapted versions of empero-ai/Qwen3.8-9B-Distill, optimized for ComfyUI inference.

This repository provides two formats:

• BF16 — Adapted Safetensors weights for ComfyUI compatibility.
• W4A8 — 4-bit weight and 8-bit activation quantization, optimized for ComfyUI inference.

Available Formats

Format	File	VRAM*	Hardware	Notes
BF16	Qwen3.8-9B-Distill-ComfyUI_bf16.safetensors	~19.3 GB	Any modern GPU	ComfyUI-adapted BF16 weights
W4A8	Qwen3.8-9B-Distill-ComfyUI_w4a8.safetensors	~10 GB	Compatible kernels	4-bit weights + 8-bit activations

*Approximate VRAM requirements depend on the inference configuration, context length, resolution and runtime implementation.

W4A8 is intended to reduce memory requirements while maintaining a good balance between model quality and inference efficiency.

Use BF16 for maximum precision and as a reference for quality comparisons.

License

These models are derived from empero-ai/Qwen3.8-9B-Distill and inherit the applicable license and usage conditions of the original model.

Please refer to the original model card for the complete license terms.

Credits

• empero-ai — Original Qwen3.8-9B-Distill model.
• silveroxides — convert_to_quant quantization toolkit (w4a8 quantization is an open PR for now, see here Hullabalo).