FLUX.2 [klein] 9B Base

FLUX.2 [klein] 9B Base is a 9-billion-parameter rectified flow transformer from Black Forest Labs, published on Hugging Face as black-forest-labs/FLUX.2-klein-base-9B. It generates images from text and edits them with multiple references. “Base” means it was trained without step or guidance distillation, which keeps it flexible for fine-tuning, LoRA training and research.

The base model is undistilled: it keeps the full sampling process and classifier-free guidance, so it needs more steps (the model card's example uses 50) but is the better starting point for fine-tunes and LoRAs. The distilled FLUX.2 [klein] 9B is tuned for speed and needs far fewer steps. Black Forest Labs says the base suits work “where control matters more than speed”.

Yes. Kenerate AI runs it in the browser as Flux 2 Klein Base 9B, with no download or setup. It costs 4.5 credits per image and 6.3 per edit, takes up to 3 reference images and offers 8 sizes up to 1536×1536. Running the weights yourself needs about 29 GB of VRAM, per the model card.

The weights are free to download from Hugging Face after you accept the licence, but you need your own high-VRAM GPU. On Kenerate the 9B Base is a paid model (4.5 credits per image from a one-time credit pack, no subscription). The free plan includes Flux 2 Klein (4B) and Flux 2 Dev if you want free FLUX.2 images.

The 9B weights are released under the FLUX Non-Commercial License, according to the Hugging Face model card. The 4B [klein] models are Apache 2.0. Read Black Forest Labs' licence before using a self-hosted 9B model commercially; images you generate on Kenerate follow Kenerate's terms of service.

It refers to an 8-bit (FP8) copy of the same 9B Base weights, which ComfyUI lists for its FLUX.2 [klein] workflows. ComfyUI Wiki puts the full checkpoint at about 18.9 GB and the FP8 file at about 9.5 GB, so FP8 fits smaller GPUs at some cost in precision. On Kenerate you don't choose a file — the hosted model runs for you.

Yes. The model card lists multi-reference editing. On Kenerate, upload up to 3 reference images to Flux 2 Klein Base 9B and describe the change and what must stay, for example “same pose, swap the background for a beach”. Each edit costs 6.3 credits.

Mixed in our test. Large Latin lettering worked — “LISBOA” on a travel poster was spelled right first try — but small text and Japanese signage came out as gibberish. For posters with lots of text, try GPT Image 2 or Qwen Image on Kenerate instead.

The Diffusers example on the Hugging Face card generates at 1024×1024 with 50 inference steps and a guidance scale of 4.0, and suggests CPU offload to save VRAM. Wiro's API example uses 30 steps and scale 4. On Kenerate the steps and guidance are set for you; you choose the prompt, size and seed.