Qwen3 4B ZIT - INT8 ConvRot

INT8 ConvRot quantization of Qwen3 4B tailored for Z-Image/Z-Image Turbo text conditioning in ComfyUI. The 8.04 GB BF16 checkpoint is reduced to approximately 4.62 GB.

Conversion

  • Tool: silveroxides/convert_to_quant
  • Format: INT8 row-wise with embedded ConvRot metadata
  • ConvRot group size: 256
  • Method: learned rounding (AdaRound) with low-memory streaming conversion
  • Quantized: 238 matrices across 34 transformer blocks (~3.43B parameters, or 85.3% of matrix parameters)

The token embedding and transformer blocks 0 and 34 remain BF16. Norms and biases also remain at their original precision.

Why this is Z-Image-specific

ComfyUI configures the Z-Image Qwen3 4B encoder with layer_idx=-2. For its 36-block transformer, this selects the hidden state produced immediately after block 34 rather than the final output from block 35.

Block 34 therefore remains BF16 because it directly produces Z-Image's conditioning representation. Block 0 is retained as a conservative input-boundary precaution. Block 35 is quantized because its output is not consumed by the Z-Image conditioning path.

This selection is tailored to Z-Image/Z-Image Turbo and is not intended as a general-purpose Qwen3 language-model quantization recipe.

Command

ctq -i <input-model>.safetensors -o qwen_3_4b_zit_int8_convrot.safetensors `
  --int8 --scaling_mode row `
  --convrot --convrot-group-size 256 `
  --comfy_quant --save-quant-metadata `
  --low-memory --device cuda `
  --exclude-layers '(^model\.embed_tokens\.weight$|^model\.layers\.(0|34)\.)' `
  --verbose NORMAL

Quantization is lossy, so outputs are not bit-identical to the original BF16 checkpoint.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Stick9190/qwen_3_4b_zit_int8_convrot

Finetuned
Qwen/Qwen3-4B
Finetuned
(1044)
this model