FLUX 3 Image Klein Brings Open Image Generation to an 8 GB Graphics Card
Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for generating and editing images. A compressed build runs on a GPU with 8 GB of memory.
Black Forest Labs has released FLUX 3 Image Klein, an image model whose weights anyone can download. It has 5.6 billion parameters, 38 percent fewer than its predecessor FLUX.2 [klein] at 9 billion. A parameter is one of the numbers a model learns during training, so fewer of them roughly means a smaller, lighter model. The practical result: an INT8 build (a compressed version that stores each number in less space) runs on an Nvidia graphics card with 8 GB of video memory. On an RTX 4070, The Decoder reports an image in under nine seconds.
The model does text-to-image, image editing, and generation from up to ten reference images. It supports native 4K output, pixel-accurate edits and bounding boxes, which let you say where in the picture an element should go. Black Forest Labs claims the highest Elo ratings among open models, beating Ideogram V4.5 and FLUX.2 [pro], and a gain of 91 Elo points over FLUX.2 [klein]. Elo is a rating borrowed from chess: people compare two images and vote, and the scores come from those votes. These are the company’s own internal results, and the article gives no absolute scores, so treat them as a promising claim rather than a verdict.
Why the size matters: most strong image generators live behind a website or an API. A model that fits on a mid-range gaming card means you can generate images on your own computer, with no per-image fee and without your prompts leaving the machine. It also comes with LoRA training from day one, a light way to teach the model a specific style or subject with a handful of your own pictures.
The licence needs a close read. The weights ship under the FLUX Kommunity License, which per the coverage is free for commercial use by companies earning under 5 million dollars a year. Larger companies need a deal. The article also lists an API price that looks like a typo, so check the official docs before budgeting anything.
What this means for you: if you are curious and own a recent gaming PC, this is a good excuse to try local image generation through ComfyUI or Hugging Face’s Diffusers library, both supported. If you have no such graphics card, the hosted playground works too. For anything commercial, read the licence first, and for most people the practical gain is simply freedom to experiment without a subscription.
Sources
Lithos-Metal Is a New Open Engine for Running AI Fast on Apple Silicon Macs
The open-source lithos-metal engine targets Qwen models on Apple's M5 chips and cites 168 tokens per second. That figure is a projection under stated assumptions, not a measured end-to-end speed.
