Black Forest Labs' FLUX 3 Klein runs on an 8 GB GPU

A strong open image model now fits on a mid-range gaming card. That changes who can run image generation in-house.

Black Forest Labs released FLUX 3 Image Klein on Oct. 15 as open weights. The company says it is a 5.6 billion parameter model for text-to-image, image editing and generation from up to 10 reference images, with native 4K output and bounding-box control over where elements go.

The memory number is the launch

BFL lists the generator weights at 11.2 GB in BF16 and 5.6 GB in FP8. An INT8 build, developed with Nvidia, runs on GeForce RTX cards with as little as 8 GB of VRAM through ComfyUI or Diffusers.

8 GB.

On speed, BFL says the step-distilled model with the INT8 build makes a 1 megapixel image in about 4 seconds on an RTX 5090 and in under 9 seconds on an RTX 4070.

There are two checkpoints: a base model for fine-tuning and a step-distilled model for fast generation. The weights are on Hugging Face, with day-one support in ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni. LoRA training works from launch through Ostris's AI Toolkit and Nvidia NeMo AutoModel, BFL says. The model is also on BFL's API and Playground.

The quality claims are BFL's own

BFL says its internal evaluations rank Klein first among open-weight models in text-to-image, multi-reference generation and editing, and above the closed Ideogram V4.5 and its own FLUX.2 [pro]. It says Klein gains 91 Elo points in text-to-image and 70 in editing over FLUX.2 [klein] 9B, with 38% fewer parameters.

Those are the company's numbers on the company's tests. Run your own prompts before you trust the ranking.

BFL says the model was red-teamed internally and evaluated by an outside partner, Cinder. This is the second open-weight release in the FLUX 3 family, after FLUX 3 Action.

Check the license before you check the pictures

The weights ship under the FLUX Kommunity License. BFL says it is free for commercial use by individuals and companies with annual revenue under $5 million, and for research and non-commercial use. Larger organizations license it for self-hosting or use the API.

That line decides what this release means for you.

Under it, you can generate product shots, ad variants and prototypes on one consumer GPU and train LoRAs on your own brand style, with no per-image bill. Over it, the free weights are for research and non-commercial use, and a commercial deployment means a license or the API.

If you already pay per image to a closed model, this is the week to run a side-by-side on your real prompts. Count the misses, not just the wins.

What to watch

Watch the independent image arenas over the next few weeks, where BFL's internal Elo claims meet public votes. Then watch what the next open release in the FLUX 3 family is.

Read the license. Then download it.