FLAN-T5-XXL (Text-Encoder Only)
The FP8 and GGUF format is distributed as a compressed ZIP file. Please unzip it using any decompression software of your choice before use, or download from Hugging Face page.
FLAN-T5-XXL is a fine-tuned version of T5-XXL v1.1, designed to improve accuracy and performance.
The original FLAN-T5-XXL model is available on Google's Hugging Face page.
When used with Flux.1, SD3.5 and HiDream, replacement for T5-XXL v1.1 to FLAN-T5-XXL offers improved prompt comprehension and enhanced image quality.
This model has been streamlined by extracting only the text encoder portion, making it optimized for image generation workflows.
Model Variants
FP32
Full precision, highest quality — largest file size and heaviest VRAM/compute footprint. Best for archival or when maximum fidelity matters more than speed.
FP16
Half precision with virtually no perceptible quality loss for most use cases. The recommended default — a solid balance of accuracy and file size.
INT8_ConvRot_HQ
Lightweight 8-bit quantization using rotation-based conversion, preserving unusually high accuracy despite the reduced size. Great if you want INT8-level savings without the usual quality trade-off.
MXFP8 / FP8_svd_scaled / FP8
⚠️ Currently not recommended. These FP8 variants are kept for compatibility/testing but show degraded quality or stability compared to the other options — use INT8_ConvRot_HQ instead.
GGUF (Q8_0 / Q6_K / Q5_K_M / Q4_K_M)
Excellent compression ratios for constrained environments. Quality scales down the quantization ladder (Q8_0 closest to full precision, Q4_K_M smallest but with more quality loss) — pick based on your available memory.
Comparisons
Tip: Upgrade CLIP-L Too
For even better results, consider pairing FLAN-T5-XXL with an upgraded CLIP-L text encoder:
LongCLIP-SAE-ViT-L-14 (ComfyUI only)
Combining FLAN-T5-XXL with an enhanced CLIP-L model can further boost image quality.
License
This model is based on Google's FLAN-T5-XXL, also licensed under Apache 2.0.
Update History
2026.7.10
Upload INT8_ConvRot_HQ
2025.6.24
Re-upload of the GGUF model, reduction in model size, and correction of metadata.
Description
FAQ
Comments (10)
interesting.
but i don´t see a huge difference compared to t5xxl_fp16.
did i miss something?
The difference in image quality is not significant and can only be distinguished using tools.
For the differences between t5xxl and flan-t5xxl, matataByy has written an article in Japanese, so please check it using your browser's translation function.
The Q8 is from the FP32?
If i'm using FP16, it makes sense trying the Q8?
The Q8 model was created from the FP32.gguf file.
If FP16 is working well for you, there’s no need to switch to Q8.
It`s a big different between fp16 and fp32?
There is a difference between FP16 and FP32 — but it's subtle.
When used together with the improved CLIP-L (FP32) encoder mentioned in the model description, you may observe better consistency and detail in generation quality.
https://www.ai-image-journey.com/2024/12/image-difference-t5xxl-clip-l.html
Personally there's no different.
the sha256 from zip will never be the same as the model so the images cant be link to this model
Civitai does not support uploading GGUF format, and FP8 format also could not be uploaded in this case, so I have compressed it into ZIP format.
The original model is uploaded to Hugging Face, where you can verify the hash value.
Example: flan_t5_xxl_TE-only_FP8.safetensors model
https://huggingface.co/easygoing0114/flan-t5-xxl-fused/blob/main/flan_t5_xxl_TE-only_FP8.safetensors
@easygoing0114 thank you. It's just that when we click the little magnifying glass in Reforge, it opens the Civitai page directly. I've seen other models in GGUF and q8 or fp8 formats; I didn't know some couldn't be uploaded. Your guides helped me a lot to get started with Flux thanks!