Step-by-step setup
1 · Check your hardware
You need an NVIDIA GeForce RTX card with at least 8 GB of VRAM for the INT8 build. Check yours in Task Manager → Performance → GPU, or with nvidia-smi.
2 · Download the weights
Get the official FLUX 3 Klein weights from the Black Forest Labs page on Hugging Face. Always download from the official organization — mirror re-uploads can be stale or tampered with. Accept the license terms when prompted.
3 · Pick an inference stack
- ComfyUI — node-based workflow, best for repeatable pipelines and ControlNet-style extras.
- diffusers — Python library, best for scripting batches and integrating into your own tools.
- MLX — Apple-silicon path if you are on a Mac.
4 · Use the INT8 quantized build
The INT8 quantization is what brings the memory floor down to 8 GB. Load the quantized checkpoint rather than full precision unless you have VRAM to spare — quality difference is minor, the memory difference is not.
5 · Read the license before selling anything
The announcement allows free in-house use under the $5M revenue threshold. If outputs feed a commercial product, confirm your case against the official license text on the model page. This site is an independent guide and not legal advice.
Specific commands follow the official model card — it is updated by Black Forest Labs as inference stacks add support. Follow it as the source of truth.