MisterNatural/Jan-v1-4B-ComfyUI

Jan-v1-4B — ComfyUI Quantizations

Quantized versions of janhq/Jan-v1-4B, optimized for ComfyUI inference.

Produced with convert_to_quant (learned rounding optimization included).

Available Formats

Format	File	VRAM*	Hardware	Notes
BF16	jan-v1-4b-ComfyUI_bf16.safetensors	~8 GB	Any modern GPU	Unmodified baseline weights
INT8 ConvRot	jan-v1-4b-ComfyUI_int8-convrot.safetensors	~4.6 GB	Any GPU	Row-wise + ConvRot (group 256/64), learned rounding
W4A8 ConvRot	Jan-v1-4B-ComfyUI_w4a8_convrot.safetensors	~2.8 GB	compatible kernels	4-bit weights + 8-bit activations, ConvRot quantization

*Approximate — actual usage depends on resolution and context.

W4A8 Convrot offers the best quality/size trade-off on consumer GPUs. For quickness or if you notice artifacts, use BF16.

License

These quantizations inherit the license of the base model janhq/Jan-v1-4B. Check the base model card for terms of use.

Credits

• janhq — original Jan-v1-4B model
• silveroxides — convert_to_quant quantization toolkit