TensorTemplar/flux2-text-encoder-fp8

FLUX.2 Text Encoder (FP8)

Combined repo for FLUX.2 text encoding with FP8 quantization (~24GB VRAM instead of ~48GB).

Components

Component	Source
FP8 Model Weights	RedHatAI/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic
Tokenizer/Processor	mistralai/Mistral-Small-3.1-24B-Instruct-2503

Usage

from transformers import AutoProcessor, Mistral3ForConditionalGeneration

model = Mistral3ForConditionalGeneration.from_pretrained(
    "TensorTemplar/flux2-text-encoder-fp8",
    local_files_only=True,  # or False for download
)
processor = AutoProcessor.from_pretrained(
    "TensorTemplar/flux2-text-encoder-fp8",
    use_fast=False,
)

Purpose

This repo exists to simplify FLUX.2 deployment by combining all necessary text encoder components into a single download. Used for extracting intermediate hidden states (layers 10/20/30) for image generation conditioning.

Attribution

• FP8 quantization by RedHatAI using llm-compressor
• Original model by Mistral AI