FLUX.2 [klein] 9B.

FLUX.2 [klein] 9B is an open-weight image model from Black Forest Labs. Specs, recommended settings, hardware and licence terms, from the developer's own pages.

Official model card

Status: Reference
Task: Text to image, Image editing
Input: Prompt, optional reference images
Output: Images
Developer: Black Forest Labs
Parameters: 9B
Licence: FLUX Non-Commercial License
Released: January 2026

What the model does.

Paraphrased from the official model card.

FLUX.2 [klein] 9B is the larger [klein] model from Black Forest Labs. Like the 4B, one model does both text-to-image and image editing, including edits that combine several reference images. It pairs a 9B flow transformer with an 8B Qwen3 text embedder and is step-distilled to 4 inference steps.

BFL also publishes an undistilled 9B Base for fine-tuning and LoRA training, and a 9B KV variant for editing that computes the reference images' attention once and reuses it across steps. All of the 9B models are under the FLUX Non-Commercial License, while the 4B models are Apache-2.0.

Reference page.

We don't currently serve FLUX.2 [klein] 9B. The specs, settings and licence terms on this page come from Black Forest Labs's own pages.

Live for images today: Z-Image Turbo (text to image) and FLUX.2 [klein] 4B (text to image). See Image generation for what the API does now.

Is FLUX.2 [klein] 9B licensed for commercial use?

Licence: FLUX Non-Commercial License.

Every model's licence

No, not under its open licence. The FLUX Non-Commercial License covers personal, research and non-production use, and commercial use needs a separate licence from Black Forest Labs. The licence lets you use the images themselves for any purpose, except training a model that competes with FLUX. Only the 4B [klein] models are Apache-2.0.

FLUX Non-Commercial License v2.1. It allows use, derivatives and distribution solely for Non-Commercial Purposes: personal research, experimentation and hobby use, testing and evaluation by companies in a non-production environment, and charitable use. Use for revenue-generating activity, in direct interactions with or with an impact on end users, or to train models for commercial use doesn't count as non-commercial. flux2 repo, LICENSE-FLUX-NON-COMMERICAL
Section 2(b): any commercial use needs a separate licence from BFL, which BFL grants at its discretion and which may carry a fee, royalty or revenue share. BFL points to bfl.ai/licensing. flux2 repo, LICENSE-FLUX-NON-COMMERICAL
Section 2(d): BFL claims no ownership of outputs, and you may use them for any purpose, including commercial ones, except as the licence prohibits. Outputs can't be used to train, fine-tune or distil a model that competes with a FLUX model. flux2 repo, LICENSE-FLUX-NON-COMMERICAL
Section 2(e): you must run content filters that prevent unlawful or infringing content (BFL's provided filters qualify) or review outputs before distributing them, and label outputs as AI-generated where the law requires it. flux2 repo, LICENSE-FLUX-NON-COMMERICAL
Section 4 also rules out military use, surveillance, biometric processing, and uses that infringe rights of publicity or "digital replica" rights. Distributing the weights or a derivative requires a copy of the licence and BFL's attribution notice. flux2 repo, LICENSE-FLUX-NON-COMMERICAL
The same licence covers [klein] 9B, 9B KV, 9B Base and FLUX.2 [dev]. Only [klein] 4B and 4B Base are Apache-2.0. flux2 repo
Trademarks: using BFL or FLUX trademarks or logos in modified versions must not cause confusion or imply sponsorship or endorsement. FLUX.2-klein-9B model card

This is our reading of the licence, not legal advice. Read the licence text before you rely on it.

FLUX.2 [klein] 9B VRAM requirements.

Published by Black Forest Labs for running the weights yourself. These are the developer's figures, not Windpaint's.

Every model's published VRAM

Published VRAM: About 29 GB
GPUs named: RTX 4090 and above
Applies to: Model card figure for the distilled 9B. FP8 and NVFP4 versions are also published.

The card says the model fits in about 29GB of VRAM and is accessible on NVIDIA RTX 4090 and above. FLUX.2-klein-9B model card
The card's diffusers example loads Flux2KleinPipeline in bfloat16 and calls pipe.enable_model_cpu_offload() to save VRAM. FLUX.2-klein-9B model card
The launch post says BFL released FP8 and NVFP4 versions of all [klein] variants, with FP8 using up to 40% less VRAM and NVFP4 up to 55% less, benchmarked on RTX 5080/5090 at 1024 × 1024. For 9B they're published as FLUX.2-klein-9b-fp8, FLUX.2-klein-9b-nvfp4, FLUX.2-klein-base-9b-fp8 and FLUX.2-klein-base-9b-nvfp4. BFL blog
The 9B pairs a 9B flow transformer with an 8B Qwen3 text embedder. FLUX.2-klein-9B model card
BFL's README recommends [klein] 4B, not 9B, for consumer GPUs such as the RTX 3090/4070. flux2 repo

FLUX.2 [klein] 9B native resolution.

What the developer's own code and examples use.

The card's example generates at 1024 × 1024. The card doesn't state a maximum. FLUX.2-klein-9B model card

Recommended settings for FLUX.2 [klein] 9B.

Steps, guidance and defaults from the developer's card and code, for running the weights yourself.

The card says the model is step-distilled to 4 inference steps. Its example uses num_inference_steps=4, guidance_scale=1.0 and 1024 × 1024. FLUX.2-klein-9B model card
BFL's README says to use the distilled models (4 steps) for production apps and real-time generation, and the undistilled Base models (50 steps) for fine-tuning, LoRA training and maximum flexibility. flux2 repo
For image editing, BFL's README points to the 9B KV variant. It computes the keys and values for the reference-image tokens once, at the first step, and reuses them for the remaining steps. The official CLI uses this path when you pick flux.2-klein-9b-kv. flux2 repo, flux2_klein_kv_cache.md

FLUX.2 [klein] 9B prompt guide.

From Black Forest Labs's own prompt guide, prompt tooling and examples.

BFL's FLUX.2 prompting guide, which is written for [pro] and [max], says FLUX.2 does not support negative prompts. Describe what you want instead: "sharp focus throughout" rather than "no blur", "empty scene" rather than "no people". BFL docs
The same guide says word order matters and FLUX.2 pays more attention to what comes first. Its suggested order: main subject, key action, critical style, essential context, secondary details. BFL docs
The same guide says to put text to render in quotation marks and give its placement, style and size, and to tie hex colours to specific objects ("The car is #FF0000"). BFL docs
The card notes that prompt following is heavily influenced by prompting style. FLUX.2-klein-9B model card

FLUX.2 [klein] 9B limitations.

As the developer states them.

Not intended or able to provide factual information. FLUX.2-klein-9B model card
It can output text, but rendered text may be inaccurate or distorted. FLUX.2-klein-9B model card
It may represent or amplify biases in its training data, may fail to match the prompt, and its prompt following depends heavily on prompting style. FLUX.2-klein-9B model card
BFL's FLUX.2 overview lists [klein] multi-reference editing as "Up to 4" images. BFL docs

Questions.

Can I call FLUX.2 [klein] 9B on Windpaint?

Not at the moment. For images today, z-image-turbo and flux-2-klein-4b are live.

How much VRAM does FLUX.2 [klein] 9B need?

The model card says it fits in about 29GB of VRAM, on an NVIDIA RTX 4090 or above. Its diffusers example also turns on CPU offload to save VRAM.

Can I use FLUX.2 [klein] 9B commercially?

Not under its open licence. The FLUX Non-Commercial License covers personal, research and non-production use only, and commercial use needs a separate licence from Black Forest Labs. The licence does let you use the images themselves for any purpose, except training a model that competes with FLUX.

Is FLUX.2 [klein] 9B open source?

No, it's open-weight. The weights are downloadable from Hugging Face once you accept the licence, but the licence is the FLUX Non-Commercial License, not an open-source one. BFL's README lists only the two 4B [klein] models as Apache-2.0.

What's the difference between FLUX.2 [klein] 9B and 4B?

The licence and the size. 4B is Apache-2.0 and 9B is non-commercial. BFL's README recommends 4B for consumer GPUs such as the RTX 3090/4070. Both are step-distilled to 4 steps and both do text-to-image plus single and multi-reference editing.

How many steps does FLUX.2 [klein] 9B use?

Four. The card says it's step-distilled to 4 inference steps, and its example runs 4 steps with a guidance scale of 1.0 at 1024x1024.

What is FLUX.2 [klein] 9B Base?

The undistilled version of [klein] 9B. It runs 50 steps, and BFL's README suggests it for research, LoRA training and maximum output diversity. It's under the same non-commercial licence.

What is FLUX.2 [klein] 9B KV?

A variant of [klein] 9B for image editing. It caches the attention keys and values of the reference images at the first step and reuses them for the rest, so references are processed once however many steps run.

Is there an FP8 version of FLUX.2 [klein] 9B?

Yes. BFL publishes FLUX.2-klein-9b-fp8, an FP8 version of the 9B weights, and its launch post says FP8 and NVFP4 versions were released for every [klein] variant.

FLUX.2 [klein] 9B.

Start yours.