⚡ Black Forest Labs · Next-Gen Rectified Flow

FLUX.2 [klein] Base 9B FP8

High-efficiency foundation image synthesis model quantized in 8-bit floating point precision (FP8) with Qwen 3 8B text conditioning supporting prompts up to 12,288 tokens.

01 · Prompt Studio & Conditioning

PROMPT (NATURAL LANGUAGE OR DETAILED DESCRIPTION)

ASPECT RATIO

INFERENCE STEPS

GUIDANCE SCALE (CFG)

SEED

MAX SEQUENCE TOKENS (QWEN 3 CONDITIONING)

128 tokens 512 tokens 12,288 tokens

CURATED EXAMPLES

Target Architecture: Rectified Flow ZeroGPU Compatible

02 · Latent Viewport

FP8 Scaled

⚡

FLUX.2 Klein Base 9B Viewport

Ready for generation

BACKBONE 9B Base Transformer

PRECISION FP8 (8-bit Float)

CONDITIONING Qwen 3 8B (12k tok)

Python Integration & Diffusers Pipeline

The complete PyTorch & ZeroGPU code for this model is embedded in app.py and ready to serve via Gradio / MCP:

from diffusers import Flux2KleinPipeline, Flux2Transformer2DModel
import torch

# Load the base 9B pipeline
pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9b-fp8",
    torch_dtype=torch.bfloat16
).to("cuda")

image = pipe(
    prompt="A colossal ancient monastery city on floating basalt islands above a sea of clouds",
    num_inference_steps=24,
    guidance_scale=3.5,
    width=1024,
    height=1024
).images[0]
image.save("flux2_output.png")