Black Forest Labs/FLUX.2-klein 9B

• License: FLUX Non-Commercial License
• Release date: 2026-01-15 (official model-family release announcement)
• Modalities: Input: text, image, multiple images; Output: image
• Original model: black-forest-labs/FLUX.2-klein-9B

Description

FLUX.2-klein 9B is a rectified-flow image model that combines text-to-image generation with single- and multi-reference editing. Its 9B flow transformer uses a separate 8B Qwen3 text encoder and is step-distilled for four-step inference, targeting interactive generation and editing. This checkpoint is the distilled 9B model, not the base or KV variant.

Run

self serve flux2-klein:9b

Here is a link to the API docs: Image API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/images/generations \
	-H 'Content-Type: application/json' \
	-d '{
		"prompt": "A red fox in a snowy forest",
		"size": "1024x1024",
		"n": 1,
		"output_format": "png"
	}' \
	-o response.json

The server returns JSON with base64-encoded image bytes in data[0].b64_json, not a hosted image URL. Decode the first result with jq and base64:

jq -r '.data[0].b64_json' response.json | base64 --decode > output.png

Engine Parameters

Served by the unmodified upstream stable-diffusion.cpp sd-server through the sd-server adapter, with one loaded pipeline per process. These are startup settings; defaults below come from the registry. Unset settings are delegated to the engine rather than assigned a default by the adapter.

Parameter	Default	What it controls
steps	40	Sampling steps; accepts integers from 1 to 200. The upstream distilled checkpoint recommends four steps, but the registry currently sets 40.
cfg_scale	1.0	Guidance scale; accepts values from 0 to 50.
sampling_method	euler	Sampling method; the adapter accepts only euler.
flash_attention	true	Enable flash attention.
offload_to_cpu	Unset (engine default)	Offload inactive model stages to CPU.
threads	Unset (engine default)	CPU workers; accepts integers from 1 to 1024.
seed	Unset (engine default)	Default seed; accepts integers from 0 to 9223372036854775807.

Override startup settings with --set <parameter>=<value>.

Use --set steps=4 to select the upstream four-step recommendation. Per-request fields such as size, n and output_format belong in the API request, not --set; a request's seed overrides the startup seed. Backend sampling overrides can be supplied through sd_cpp_extra_args.sample_params; see the Image API for the backend-specific request format. Omitting overrides keeps the startup sampling defaults.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry. The pipeline requires all three files listed, including the separate encoder and safetensors VAE.

q4 (default)

File	Role	Size	SHA-256	Origin
flux-2-klein-9b-Q4_K_M.gguf	model	5.50 GiB (5909829920 bytes)	5489463ed96056b0bb5472abb5d1bba7055e48d574e37877acb43b407465e26f	unsloth/FLUX.2-klein-9B-GGUF
split_files/vae/flux2-vae.safetensors	vae	320.6 MiB (336211292 bytes)	868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3	Comfy-Org/flux2-klein
Qwen3-8B-Q4_K_M.gguf	text-encoder	4.68 GiB (5027784512 bytes)	120307ba529eb2439d6c430d94104dabd578497bc7bfe7e322b5d9933b449bd4	unsloth/Qwen3-8B-GGUF

Total download: 10.50 GiB (11273825724 bytes).

Notes

• The 9B checkpoint uses the FLUX Non-Commercial License, not the 4B model's Apache-2.0 license. Review all license and acceptable-use restrictions before deployment; the original card requires filters or manual review under the license.
• The API exposes POST /v1/images/generations and multipart POST /v1/images/edits, with at most eight reference images. See the Image API for uploads and remote reference URLs.
• Rendered text may be inaccurate, prompts may not be followed, and outputs may reflect training-data biases. The model is not a source of factual information.
• Download size is not the total runtime memory requirement. Upstream latency claims do not establish performance for this GGUF pipeline.
• Native inference, editing and GPU performance still need verification on target hardware; registration and mocked protocol tests are not an inference benchmark.

Quality benchmarks

Arena Text-to-Image: #62 (1070.29)

flux-2-klein-9b

Upstream model; not this local quantization

Published 2026-10-07; 150782 votes

Checked 2026-10-13

Arena Image Edit: #42 (1224.74)

flux-2-klein-9b

Upstream model; not this local quantization

Published 2026-10-06; 540742 votes

Checked 2026-10-13

Benchmarks

No submitted benchmark reports yet.