Image generation and editing models often demand substantial computing resources. Black Forest Labs’ new FLUX 3 Image Klein is designed to make those tasks more accessible: its INT8 build can run on NVIDIA GPUs with 8 GB of VRAM, and the company says it can generate an image in under 9 seconds on an RTX 4070.
One model for creating and editing images
FLUX 3 Image Klein is a 5.6-billion-parameter open-weights model. It supports text-to-image generation, image editing, and generation guided by up to ten reference images. That combination gives users several ways to steer an image, from starting with a written description to working from existing visual examples.
The model also supports native 4K output and pixel-accurate editing. For placement tasks, it can use bounding boxes to position elements. These features point to a workflow where users can specify both what an image should contain and where particular elements should appear, then make precise changes to the result.
Performance claims and local hardware
Black Forest Labs says FLUX 3 Image Klein achieved the highest Elo ratings among open models in the company’s internal testing. The company also says it beat closed competitors Ideogram V4.5 and its own FLUX.2 [pro]. These are results attributed to the company’s testing, rather than an independent assessment described in the source.
Compared with FLUX.2 [klein], the new model gains 91 Elo points while using 38 percent fewer parameters. FLUX.2 [klein] has 9 billion parameters. The comparison suggests the new release is intended to improve on its predecessor while using a smaller model, though the source does not provide details about the test setup.
For people running models locally, the INT8 build is a key practical detail. It allows the model to run on NVIDIA GPUs with only 8 GB of VRAM, and Black Forest Labs reports image generation in under 9 seconds on an RTX 4070. That gives a concrete example of the hardware and speed the company associates with the release.
Ways to access the model
Users can try or integrate FLUX 3 Image Klein through several routes. It is available via the BFL API, Playground, Hugging Face Spaces, and locally through ComfyUI and Diffusers. The source lists the API price as $0.009 per 1,000 images.
Those options serve different needs: a hosted service offers access without a local setup, while ComfyUI and Diffusers provide ways to use the model on a local machine. The source does not specify the requirements or feature differences for each option, so users will need to choose based on their workflow.
Licensing and customization
The weights are distributed under the FLUX Kommunity License. The source says commercial use is free for companies making under $5 million annually. That condition matters for businesses evaluating whether the weights can be used in their products or production workflows.
LoRA training is supported from day one through Ostris AI Toolkit and NVIDIA NeMo AutoModel. This makes customization part of the initial release: users can adapt the model with those tools rather than relying only on its default behavior.
FLUX 3 Image Klein brings image creation, editing, reference-guided generation, and model customization into one release. Its 8 GB VRAM option and multiple access paths may make experimentation easier for people without high-memory GPUs, while its license terms and attributed performance claims are important context for companies considering broader use.