sjoe1244/Qwen3-8B-FP8-native-for-FLUX2-ComfyUI

Qwen3-8B Native FP8 Text Encoder for FLUX.2 ComfyUI

This repository contains a ComfyUI-native FP8 conversion of Qwen/Qwen3-8B intended for FLUX.2 / Flux 2 text-encoder workflows.

It was converted from the BF16 Qwen/Qwen3-8B weights, not from HF compressed FP8 shards. In practice this avoids the extra compressed-FP8 scale conversion step and loaded cleanly in ComfyUI's CLIPLoader with type=flux2.

File

• Qwen3-8B-FP8-native-bf16.safetensors
• Size: 8,191,089,400 bytes
• SHA256: 2f3ef1c5dac47bc7f423b8b306c4ad5ddfdface9da7efa3f250da9d8b440af84

Conversion

• Source model: Qwen/Qwen3-8B
• Floating 2D tensors were cast to torch.float8_e4m3fn.
• Other tensors, such as norms and scalar parameters, were kept in BF16.
• Output is a single .safetensors file for easy placement in ComfyUI/models/clip.

The included convert_qwen3_bf16_to_native_fp8.py script documents the exact local conversion pattern used for this upload.

Usage

Place the safetensors file in one of:

• ComfyUI/models/clip/
• ComfyUI/models/text_encoders/

Then use CLIPLoader with type=flux2 and select Qwen3-8B-FP8-native-bf16.safetensors.

This is not a standalone chat model. It is intended as a text encoder for image generation workflows, especially FLUX.2 workflows that expect Qwen3 8B.

FLUX.2 note

This upload contains only Qwen-derived weights under the Apache-2.0 license. It does not include Black Forest Labs FLUX.2 weights, VAE files, or BFL text encoder files. To use official BFL FLUX.2 assets, accept and download them from the official BFL repository according to its license terms:

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B