Update model card for corrected NVFP4 build

README.md

@@ -15,24 +15,31 @@ tags:
 
 ComfyUI-native NVFP4 safetensors conversion of `Minthy/ToriiGate-0.5`.
 
-## File
+## Files
 
-`toriigate_0.5_qwen35_4b_nvfp4_comfyui.safetensors`
+- `toriigate_0.5_qwen35_4b_nvfp4_comfyui.safetensors`
+- `toriigate_0.5_qwen35_4b_nvfp4_calibrated_comfyui.safetensors`
 
-Place it in:
+Place the file in:
 
 `ComfyUI/models/text_encoders/`
 
-Then use ComfyUI's Qwen3.5 / TextGenerate workflow.
+Then use ComfyUI `CLIPLoader + TextGenerate`.
 
-## Notes
+## Important Fix
 
-This is not a Hugging Face compressed-tensors directory. It is a single ComfyUI-native safetensors file with `_quantization_metadata`.
+`model.visual.pos_embed.weight` is intentionally kept full precision.
 
-Quantized layers: 348
-Copied tensors: 376
-Output size: 3.567 GiB
-Build time: 0.2 minutes
+It is an Embedding table, not a Linear weight. ComfyUI expects it as `[2304, 1024]`.
+Packing it as NVFP4 would turn it into `[2304, 512]` and break loading.
+
+## Build
+
+Quantized layers: 347
+Copied tensors: 377
+Base output size: 3.570 GiB
+Calibrated output size: 3.570 GiB
+Activation input scales added: 347
 GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
 Torch: 2.12.0+cu130
 CUDA: 13.0