Can I run or fine-tune FLUX.2 [klein] locally?

Yes. FLUX.2 [klein] is available as open weights, so you can run it on your own hardware and fine-tune it on your own data. This is what sets [klein] apart from the hosted-only FLUX.2 models.

Licensing

FLUX.2 [klein] 4B is released under Apache 2.0, so you can use it commercially with no fees or approvals. FLUX.2 [klein] 9B is released under the FLUX Non-Commercial License; commercial use of the local weights requires a license, or you can call it through the API where commercial rights are included.

Downloading the weights

Get the models from Hugging Face: klein 4B Base and klein 9B Base. The Base variants are undistilled, which makes them the right starting point for fine-tuning.

Hardware

klein 4B trains on an NVIDIA GPU with 12GB VRAM (for example an RTX 3060 12GB) and 32GB system RAM. klein 9B needs around 22GB VRAM (for example an RTX 3090 or 4090) and 64GB system RAM.

Fine-tuning tools

Two open-source frameworks cover most workflows: ai-toolkit (all-in-one suite with GUI and CLI) and Hugging Face Diffusers (DreamBooth and LoRA examples). A LoRA is lightweight, trains in roughly 1 to 3 hours on a consumer GPU, and is easy to share.

Serving a trained LoRA without a local GPU

You can upload your .safetensors in the dashboard under Customization, then Finetunes, and call the fine-tuned endpoint with the resulting finetune_id.

For the full walkthrough, see the FLUX.2 [klein] Training guide at docs.bfl.ai.