AxForge
Sign in Get started

Model reference · open weights

FLUX.2-klein-kv

Image black-forest-labs Image edit 2 builds Non-commercial 48k dl/mo

FLUX.2-klein-kv is an open-weight image model from Black Forest Labs. FLUX.2-klein-9b-kv (BF16) weighs 34.7 GB; the smallest configuration that runs it is RTX 5090 32 GB.

  • FLUX.2-klein-kv is a 9B parameter image-to-image model developed by Black Forest Labs that supports KV-cache optimization for accelerated multi-reference editing.
  • It is step-distilled to 4 inference steps, uses an 8B Qwen3 text embedder, and is designed for interactive applications where the same reference images are used across multiple generations.
  • The model requires approximately 29GB of VRAM, is available for non-commercial use, and supports English.

Summary of the black-forest-labs/FLUX.2-klein-9b-kv-fp8 model card, 2026-10-01. The estimate below is for another build of the family.

Hugging Face ↗GitHub ↗

What it is

Released byBlack Forest Labs
Released2026-03-09
VRAM34.7 GB for the weights

What it runs on

Memory and cards for FLUX.2-klein-9b-kv (BF16)

34.7 GBweights, file size
18.2 GBbiggest part
5.0 GBworking memory, one 1024² image
537 MBruntime overhead
CardOne imageMemory
RTX 3060 12 GB … RTX 4090 24 GBdoes not fit
RTX 5090 32 GBfits (encoders offloaded)31.0 GB
L40S 48 GBtight44.0 GB
A100 80 GBfits78.2 GB
H100 80 GBfits78.1 GB
RTX PRO 6000 Blackwell 96 GBfits93.8 GB
DGX Spark (GB10) 128 GB unifiedfits107 GB
H200 141 GBfits138 GB
B200 180 GBfits176 GB

Builds

Sizes, precisions & builds

BuildParamsPrecisionWeightsSmallest setup
FLUX.2-klein-9b-kv-fp8 ↗
the transformer alone, to use with the text encoders and VAE		 —FP8 9.8 GBRTX 4060 Ti 16 GB
FLUX.2-klein-9b-kv (above) ↗ 9.1BBF16 34.7 GBRTX 5090 32 GB

From the model card

What Black Forest Labs says about FLUX.2-klein-kv

Read the model card

FLUX.2 [klein] 9B-KV is an optimized variant of FLUX.2 [klein] 9B with KV-cache support for accelerated multi-reference editing. This variant caches key-value pairs from reference images during the first denoising step, eliminating redundant computation in subsequent steps for significantly faster multi-image editing workflows.

For more information about FLUX.2 [klein], please read our blog post.

Key Features

  1. KV-Cache Optimization: Reference image KV pairs are computed once and cached, reducing computation and speeding up inference by up to 2.5 times for multi-reference editing tasks.
  2. All capabilities of FLUX.2 [klein] 9B: sub-second generation, text-to-image, and multi-reference editing in a single unified model.
  3. Ideal for interactive applications and real-time editing pipelines where the same reference images are used across multiple generations.
  4. 9B flow model with 8B Qwen3 text embedder, step-distilled to 4 inference steps.
  5. Available for non-commercial use.

How KV-Caching Works

In standard image editing, reference image tokens are processed at every denoising step. With KV-caching:

  • Step 0: Full forward pass processes reference tokens and extracts their key-value pairs into a cache.
  • Steps 1-3: Cached KV pairs are reused, skipping redundant reference token computation.

This is particularly beneficial when:

  • Editing with multiple reference images
  • Generating variations with the same references
  • Building interactive editing applications

Usage

We provide a reference implementation in our GitHub repository.

API Endpoints

FLUX.2 [klein] 9B-KV is available via the BFL API at bfl.ai.

Limitations

  • This model is not intended or able to provide factual information.
  • While the model can output text, text rendered may be inaccurate or subject to distortion.
  • As a statistical model, this checkpoint may represent or amplify biases observed in the training data.
  • The model may fail to generate output that matches the prompts.
  • Prompt following is heavily influenced by the prompting style.

Out-of-Scope Use

This model and its derivatives may not be used outside the scope of the license, including for unlawful, fraudulent, defamatory, abusive, or otherwise violative purposes as further explained in our Usage Policies.

Hardware

The FLUX.2 [klein] 9B-KV model fits in ~29GB VRAM and is accessible on NVIDIA RTX 5090 and above.

Responsible AI Development

Black Forest Labs is committed to responsible model development and deployment. Prior to releasing FLUX.2 [klein] 9B-KV, we evaluated and mitigated a number of risks, including child sexual abuse material (CSAM) and nonconsensual intimate imagery (NCII). For detailed information about our mitigations, evaluation processes, content provenance features, and policies, please see our post: Capable, Open, and Safe: Combating AI Misuse.

To report safety concerns, contact safety@blackforestlabs.ai.

License

This model falls under the FLUX Non-Commercial License.

Trademarks & IP

This project may contain trademarks or logos for projects, products, or services. Use of Black Forest Labs and FLUX trademarks or logos in modified versions of this project must not cause confusion or imply sponsorship or endorsement. Any use of third-party trademarks, intellectual property or logos are subject to those third-party's policies.

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

How it works

How image models work

Text promptwhat to makeText encoderunderstands itDiffusion stepsdenoise to pixelsImagePNG / JPEGA diffusion model starts from noise and denoises it, guided by your prompt, into a finished image.

Running it yourself

Run it on a rented GPU

Rent a machine by the hour. ComfyUI is installed on it. Open ComfyUI through the tunnel, then Templates → Flux.2 [Klein] 9B: Image Edit, and choose this model's file in its model loader. 

# on your rented machine (the ssh line is on its page in the console)
# get REPO FILE FOLDER: one file into /workspace/models/FOLDER, where ComfyUI loads it from
get() { hf download "$1" "$2" --local-dir /workspace/hf-files && mkdir -p "/workspace/models/$3" && mv "/workspace/hf-files/$2" "/workspace/models/$3/$4"; }

# the model (9.1 GB)
get black-forest-labs/FLUX.2-klein-9b-kv-fp8 flux-2-klein-9b-kv-fp8.safetensors diffusion_models
# what the template also loads “Flux.2 [Klein] 9B: Image Edit” (8.3 GB)
get Comfy-Org/flux2-klein-9B split_files/text_encoders/qwen_3_8b_fp8mixed.safetensors text_encoders
get black-forest-labs/FLUX.2-small-decoder full_encoder_small_decoder.safetensors vae

start-comfyui
Renting a GPU: connect, tunnels, ComfyUI 
# on your computer, in a second terminal: ComfyUI in your browser at http://localhost:8188
# HOST and PORT are your machine's, from its page in the console
ssh -L 8188:localhost:8188 dev@HOST -p PORT
© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms