jameriles/Qwen-Image-2.1-Hybrid-W4A8-INT8-ConvRot

Qwen Image 2.1 — Hybrid W4A8 / INT8 ConvRot for ComfyUI

Community quantization pack of the official Qwen Image 2.1 transformer and its Qwen3-VL 8B text encoder, prepared by Agustín Meriles (ameriles) for ComfyUI.

The pack combines asymmetric W4A8 for most linear layers with INT8 tensorwise + ConvRot for the layer families that the reference Q4_K_M models preserve at higher precision. It reduces the two BF16 source files from about 29.6 GiB to about 11.1 GiB while retaining selected sensitive projections in INT8.

Files

File	Destination	Size
qwen_image_2.1_hybrid_w4a8_int8_convrot.safetensors	ComfyUI/models/diffusion_models/	4.79 GiB
qwen3vl_8b_hybrid_w4a8_int8_convrot.safetensors	ComfyUI/models/text_encoders/	6.28 GiB
*.quant.json	Optional provenance sidecars; keep beside the matching model if desired	—
workflows/Qwen_Image_2.1_Hybrid_W4A8_INT8_2GPU.json	Example two-GPU ComfyUI workflow	—

The Qwen Image 2.1 VAE is not redistributed here. Download qwen_image_2.1_vae_bf16.safetensors from the official ComfyUI model release.

Quantization recipe

Transformer

• 128 Linear layers: asymmetric W4A8
• 64 Linear layers: INT8 tensorwise + ConvRot
• 73 other tensors: preserved from the BF16 source
• INT8 families in all 32 transformer blocks: attn.to_v and img_mlp.out

This selector reproduces the precision split of the reference Q4_K_M mapping: Q4_K families become W4A8 and Q6_K families become INT8 ConvRot.

Qwen3-VL 8B text encoder

• 216 Linear layers: asymmetric W4A8
• 38 tensors: INT8 tensorwise + ConvRot
• 496 other tensors: preserved from the BF16 source
• INT8 block families: self_attn.v_proj and mlp.down_proj in the blocks identified as Q6_K by the official Q4_K_M reference
• model.embed_tokens and lm_head are kept in INT8, at higher precision than their Q4_K/Q6_K GGUF counterparts

The exact selectors, expected layer counts and validation tests are published in ameriles/comfy-quant-bench.

Compatibility

• ComfyUI-native prequantized Safetensors format
• converted with Comfy Kitchen 0.2.35
• tested with PyTorch 2.9.1+rocm6.4
• runtime-tested on 2 × AMD Radeon RX 6700 XT 12 GB (RDNA2)

These formats are not exclusive to Radeon or RDNA2. Actual kernel availability, memory use and performance depend on the installed ComfyUI, Comfy Kitchen, PyTorch backend and GPU. This release has only been performance-tested on the ROCm setup above.

Setup

1. Install or update ComfyUI to a version with native mixed-quantized Safetensors support and Comfy Kitchen.
2. Copy the transformer to ComfyUI/models/diffusion_models/.
3. Copy the text encoder to ComfyUI/models/text_encoders/.
4. Download the official qwen_image_2.1_vae_bf16.safetensors into ComfyUI/models/vae/.
5. Restart ComfyUI and select the new files in your loaders.

The included reference workflow uses ComfyUI-MultiGPU and the allocation validated on the test system:

• transformer: cuda:1,100%
• text encoder: cuda:2,100%
• VAE: cuda:2

The model files themselves do not require this particular allocation or the MultiGPU extension.

Measured result

One validated 1024 × 1024 image-edit run with two reference images and 20 sampling steps completed in 484.16 seconds (about 8:04) after moving the complete transformer and text encoder onto separate 12 GB GPUs. This is a setup-specific observation rather than a universal benchmark: input resolution, prompt, workflow, thermal state, PCIe topology and software versions all affect end-to-end time.

Quantization provenance

Transformer source: Qwen/Qwen-Image-2.1, qwen_image_2.1_bf16.safetensors

• preset: qwen-image21-q4km-sensitive
• output: 5,142,880,624 bytes
• sampled INT8-vs-BF16 relative-error median: 0.0099988
• sampled relative-error maximum: 0.0101992

Text-encoder source: official Qwen3-VL 8B BF16 encoder used by the ComfyUI Qwen Image 2.1 workflow

• preset: qwen3vl8b-q4km-sensitive
• output: 6,741,467,072 bytes
• sampled INT8-vs-BF16 relative-error median: 0.0097945
• sampled relative-error maximum: 0.0101816

Both output files passed structural validation and tensor-by-tensor comparison against their BF16 sources. See the JSON sidecars for machine-readable details and SHA256SUMS for integrity hashes.

Notes

• These files target ComfyUI and are not drop-in replacements for standard Diffusers or Transformers checkpoints.
• Visual quality and prompt adherence were validated through repeated image-edit runs, not a formal perceptual benchmark.
• A recent ComfyUI/Comfy Kitchen build is required to load the embedded mixed-quantization metadata.

License and modification notice

This is a quantized derivative of Qwen Image 2.1 and is distributed under the Qwen Research License. Read the complete included LICENSE before downloading or using the files.

The original BF16 files were modified by quantization in October 2026 by Agustín Meriles. Qwen did not produce or endorse this quantization.

Credits

• Base model: Qwen/Qwen-Image-2.1
• Official ComfyUI release: Comfy-Org/Qwen-Image-2.1
• Quantization tooling and research base: JoaoZaokk/comfy-quant-bench
• Quantization, testing and release: Agustín Meriles / ameriles