Qwen-Image-2.1 · Text Encoder Guide · Updated October 2026

`qwen3vl_8b_int8_convrot.safetensors`
Download & Setup Guide

The 9.35 GB INT8-quantized Qwen3-VL 8B text encoder for Qwen-Image-2.1 in ComfyUI. What it is, where to download it, where it goes, and which quantization you actually need for your GPU.

What is this file?

qwen3vl_8b_int8_convrot.safetensors is the text encoder component of the Qwen-Image-2.1 image generation pipeline. Qwen-Image-2.1 is a 7B unified text-to-image / image-editing model, but the full pipeline is 16.2B parameters — more than half of them live in this Qwen3-VL 8B vision-language model, which encodes your prompt (and reference images) before the diffusion transformer does its work.

"INT8 ConvRot" means the weights are quantized to 8-bit integers with a rotation-based transform. The practical effect: almost the quality of full precision at roughly half the VRAM.

File specs

File	`qwen3vl_8b_int8_convrot.safetensors`
Size	9.35 GB (9,350,798,360 bytes)
Precision	INT8 (ConvRot quantized)
Model	Qwen3-VL 8B (8.767B params)
Goes in	`ComfyUI/models/text_encoders/`
Used with	Qwen-Image-2.1 / Qwen-Image-2.1-Turbo (ComfyUI)

Which text encoder should you download?

There are three official variants. Pick by your VRAM:

File	Size	VRAM need	Best for
`qwen3vl_8b_bf16.safetensors`	17.53 GB	Highest	Max quality, 32GB+ cards
`qwen3vl_8b_int8_convrot.safetensors`	9.35 GB	Medium	Best balance — RTX 4090 24GB
`qwen3vl_8b_w4a8.safetensors`	5.88 GB	Lowest	Limited VRAM (16GB cards)

On a 24 GB RTX 4090, the INT8 combo (int8 transformer + this int8 encoder + bf16 VAE) keeps all three models resident at 1 megapixel output. With the BF16 encoder, weights get paged on and off the card and generation slows down.

Download

Official repack from Comfy-Org. Install huggingface_hub first (pip install -U huggingface_hub), then from your ComfyUI root:

hf download Comfy-Org/Qwen-Image-2.1 \
  text_encoders/qwen3vl_8b_int8_convrot.safetensors \
  --local-dir models

Or grab the full recommended INT8 set in one go:

hf download Comfy-Org/Qwen-Image-2.1 \
  diffusion_models/qwen_image_2.1_int8_convrot.safetensors \
  text_encoders/qwen3vl_8b_int8_convrot.safetensors \
  vae/qwen_image_2.1_vae_bf16.safetensors \
  --local-dir models

Setup in ComfyUI

Place the file in ComfyUI/models/text_encoders/ (exact folder — the loader looks there).
Put the diffusion model in models/diffusion_models/ and the VAE in models/vae/.
Update ComfyUI to the latest version (native Qwen-Image-2.1 support since launch day), then restart.
Load the official Qwen-Image-2.1 text-to-image or edit workflow template — the CLIP loader picks up the encoder automatically.

CUDA out of memory? Drop to 1024px output, keep batch size at 1, or switch the encoder to the W4A8 variant. Confirm the file is in text_encoders/ and not clip/ — the most common setup mistake.

FAQ

INT8 vs BF16 — how much quality do I lose?

ConvRot INT8 is one of the better quantization schemes for this architecture. Most users report no visible difference at 1MP; differences show up mainly in fine text rendering at 2K. If VRAM allows, BF16 is still the reference.

Can I use it with the Turbo checkpoint?

Yes. Qwen-Image-2.1-Turbo uses the same text encoder — pair qwen_image_2.1_turbo_bf16.safetensors (or the int8 transformer) with this encoder file.

Can I use it commercially?

No — not without a separate license. Qwen-Image-2.1 is released under the Qwen Research License Agreement, which limits use to non-commercial research/evaluation. Paid services, client work, or monetized publishing need a commercial license from the rights holder.

GGUF instead?

Community GGUF quants (Q4_K_M recommended) exist via ComfyUI-GGUF, but the safetensors INT8 files above are the official Comfy-Org repack and the best-supported path in stock ComfyUI.