FLUX 3 Image Klein

TL;DR Highlight

A 5.6B open-weights image model that runs on 8GB VRAM and claims top rank in BFL's evals

Who Should Read

Backend or full-stack developers adding image generation or editing via their own GPUs or an API

Core Mechanics

5.6B open-weights image model released
- It supports text-to-image, image editing, and multi-reference generation with up to 10 reference images.
- It offers native 4K generation and bounding boxes to control placement and confine edits.
Claims top open model in internal evals
- In BFL's internal evaluations it ranks first in text-to-image, multi-reference, and editing ahead of larger open models.
- Versus FLUX.2 [klein] 9B it gains 91 Elo in text-to-image and 70 in editing with 38% fewer parameters.
Runs on 8GB of VRAM
- Generator weights are 11.2 GB in BF16 and 5.6 GB in FP8.
- The INT8 build runs on RTX GPUs with as little as 8GB VRAM via ComfyUI or Diffusers.
Speed numbers for the distilled model
- With the 16-step distilled model and INT8, a 1MP image takes about 4s on an RTX 5090 and under 9s on an RTX 4070.
- On NVIDIA GB300 with NVFP4, a 1MP image takes under 0.6s.

Community Reactions

Reddit rumor of weak people images
- A comment said Reddit users claim it is bad at people, apparently because near-NSFW content was removed from training data.
Good editing but weak fine text details
- One user found editing very solid but was disappointed by fine text details.
- They had not seen the reported body-horror outputs themselves.
- They suggested MOSS-Image for text-to-image instead.
Worse than FLUX 2 Klein, some say
- One commenter found it very disappointing compared to FLUX 2 Klein.
- Another said it is very fast on OpenRouter but clearly worse than the regular FLUX 3 model.
Reaction to pricing
- A commenter called $0.009 very competitive if quality holds up.
- They contrasted it with Krea 3, released the day before, starting at $0.06 per image.

How to Apply

Run five of your prompts through the API
- Today, generate five real prompts from your service with flux-3-image-klein on the BFL API.
- Include people and in-image text to check whether the weaknesses from the comments reproduce.
Run the INT8 build locally on an 8GB GPU
- Download the INT8 build and measure 1MP generation time on your GPU via ComfyUI or Diffusers.
- If data cannot leave your infrastructure, compare it as a self-hosting candidate.
Use bounding boxes to limit edits
- Constrain edit regions with bounding boxes so unintended areas do not change.
- Use pixel restoration in the API to preserve unedited regions exactly.
Check the license, then try a LoRA
- First confirm whether you are under $5M annual revenue, otherwise choose self-hosted licensing or the API.
- For a brand style, train a LoRA on the base model with AI Toolkit.

Try It

copy & run

# Paste into a terminal (only the install command, based on the stated Hugging Face Diffusers support)

pip install -U diffusers transformers accelerate torch

# The model ID is not given in the article, so no runnable code is included.
# 1) Find the distilled model ID via the 'Weights on Hugging Face' link on the article page.
# 2) Copy and run the Diffusers usage example from that model card.

# Next to try: generate the same prompt with the INT8 and BF16 builds and compare time and quality.

Terminology

Open weightsThe trained model files are released so you can download and run them. This differs from open source, which also releases all the code.

Step distilledA model compressed to do in few steps what a larger model does in many. You trade a little quality for speed.

EloA chess-style score from head-to-head comparisons of outputs by people or models. A higher number means it wins more often.

Pareto frontierThe boundary of best options on two trade-off axes like speed and quality, where nothing is better on both.

LoRAA lightweight fine-tuning method that trains small add-on weights instead of the whole model to teach a style or subject.