Qwen3.5 (quantized)

Quantized versions of text_encoders/qwen3.5_4b_bf16.safetensors from Comfy-Org/Qwen3.5. Nothing else is changed.

File	Format
qwen3.5_4b_int8_convrot	int8 ConvRot
qwen3.5_4b_fp8_scaled	fp8 e4m3fn

The int8 ConvRot file is probably the better choice: it stays closer to bf16.

Only the language-model linear layers are quantized, in ComfyUI's comfy_quant format (int8: ConvRot group 256 with per-row scales; fp8: per-tensor scale). Embeddings, norms, the vision encoder and the MTP head stay bf16.

Put the file in ComfyUI/models/text_encoders.