FLUX 3 Image Klein: open weights, 5.6B parameters

Today we’re releasing FLUX 3 Image Klein, a 5.6B parameter open weights mannequin for text-to-image technology, picture enhancing and multi-reference technology with as much as 10 reference photographs. It shares core capabilities with the complete FLUX 3 Image mannequin: native 4K technology, pixel-perfect enhancing, and bounding containers that management the place components are positioned or confine edits to an outlined space.

FLUX 3 Image Klein builds on FLUX 3’s shared learnings throughout photographs, video and audio. Images contribute spatial element, video provides movement and bodily dynamics, and audio supplies details about occasions and interactions. Together, these indicators assist the mannequin generate photographs that really feel grounded in the actual globe.

The launch features a base mannequin for fine-tuning and a step distilled mannequin for low-latency technology, together with FP8, INT8 and NVFP4 builds developed with NVIDIA help, below the FLUX Kommunity License.

Try it in the BFL API · Download the weights

Leading open weights high quality with fewer parameters

FLUX 3 Image Klein was additionally optimized for client GPUs. Fewer parameters make it sensible to run domestically and cheaper to adapt, and groups with knowledge sovereignty or native deployment necessities can consider and adapt the mannequin on their very own roads. This can also be key for on-device picture enhancing, permitting shoppers to edit photographs immediately on their favourite laptops. The generator weights require 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 construct, FLUX 3 Image Klein runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, via ComfyUI or Diffusers, extending what FLUX.2 [klein] began to even smaller {hardware}. With the distilled mannequin and the INT8 construct, it generates a 1MP picture in about 4 seconds on an NVIDIA GeForce RTX 5090 and below 9 seconds on an RTX 4070 via ComfyUI.

The base checkpoint is constructed for fine-tuning: Diffusers and Ostris’s AI Toolkit help LoRA coaching from day one, and NVIDIA’s NeMo AutoModel ships recipes that work immediately from the launched checkpoints.

For even sooner manufacturing workloads that want decrease latency, we additionally work with groups on particular customizations that get you there. Get in touch.

On the velocity–high quality frontier

FLUX 3 Image Klein additionally extends the velocity–high quality Pareto frontier. Through the BFL API, it delivers decrease latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5 and Krea 2, and the best ranking at its latency and above. In the charts under, the violet line traces the best ranking obtainable at every latency and the grey line reveals the earlier frontier.

Latency is measured finish to finish, from request submission to the completed 1 MP picture, together with queueing. FLUX 3 Image Klein was timed on our servers. Other fashions have been timed from the consumer, which incorporates polling and community time. Krea 2 was timed on its Medium tier; its Large tier took 31.1 seconds for text-to-image.

Get began

FLUX 3 Image Klein is accessible at this time as open weights and thru the BFL API.

Try it

• BFL API: the distilled mannequin is accessible as flux-3-image-klein, with immediate upsampling and pixel restoration, which preserves unedited areas of a picture precisely. Pricing begins at $0.009 per 1K picture.
• Playground or Hugging Face.

Build with it

• Weights on Hugging Face: base and distilled.
• Reference inference code, together with a Gradio demo for bounding-box management.
• Try it out on Hugging Face Spaces.
• Day-one help in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni.
• LoRA coaching with Ostris’s AI Toolkit and recipes in NVIDIA NeMo AutoModel.
• Trained on NVIDIA GB200 NVL72 and optimized for NVIDIA’s accelerated computing platform, serving via NVIDIA TensorRT LLM, vLLM-Omni and SGLang Diffusion.
• On NVIDIA GB300, the distilled mannequin generates a 1MP picture in below 0.6 seconds in NVFP4. Serving a LoRA provides round 5% to technology time.

License

The weights can be found below the FLUX Kommunity License, which permits free business use for people and corporations with annual turnover below $5M, alongside analysis and non-commercial use. Larger organizations can deploy commercially via our self-hosted licensing choices or API.

Safety

In an effort to make AI expertise each open and secure, we implement a number of safeguards all through the mannequin improvement lifecycle to fight misuse. These embody dataset filters throughout pre-training to restrict publicity to dangerous content material, and focused fine-tuning throughout post-training to suppress dangerous behaviors, making it progressively tougher to elicit dangerous content material. Ahead of the discharge, the FLUX 3 Image Klein checkpoints have been rigorously pink teamed internally and evaluated by our exterior security associate Cinder, demonstrating considerably fewer vulnerabilities than different well-liked open weights fashions. More on our security methodology is accessible here.

FLUX 3 Image Klein is the second open weights launch from the FLUX 3 household, following FLUX 3 Action. More will observe.

Learn extra

• FLUX 3: Real World Models, the analysis behind the FLUX 3 household
• FLUX 3 Action, our open weights 7B mannequin for robotic management

Citation

If you employ FLUX 3 Image Klein in your work, please cite:

@misc{bfl2026flux3imageklein,
  creator       = {{Black Forest Labs}},
  title        = {{FLUX} 3 Image Klein},
  12 months         = {2026},
  url          = {https://bfl.ai/blog/flux-3-image-klein},
}