What it is
FLUX 3 Image Klein was released on 15 October 2026. It is a 5.6-billion-parameter image generator, with a separate Qwen3-VL-4B text encoder, that handles text-to-image, editing and generation from up to 10 reference images in one model. It inherits native 4K output and bounding-box control from the larger FLUX 3 Image.
Two checkpoints are published: a 16-step distilled model for fast output and a 50-step base model for fine-tuning and LoRA training, each in FP8, INT8 and NVFP4 builds. BFL says the INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM. The model card says not to run it without a prompt upsampler; BFL's own is a separate 9B model that adds about 28GB of GPU memory.
The weights ship under the FLUX Kommunity Licence: free commercial output for individuals and companies under US$5 million in annual revenue, counting affiliates, plus research and non-commercial use. Larger organisations need a paid self-hosting licence or the API. It succeeds the FLUX.2 [klein] models from January 2026, of which the 4B version is under the more permissive Apache 2.0.
Our review
Klein's case rests on price and licence, because its quality is still unverified. BFL's internal Elo tests put it 91 points above FLUX.2 [klein] 9B in text-to-image and 70 in editing, and ahead of Qwen-Image 2.1, Krea 2, Ideogram V4.5 and FLUX.2 [pro]. The post gives no absolute scores, VentureBeat notes the charts omit prompt sets, sample sizes and uncertainty, and a week after launch Artificial Analysis's text-to-image board still does not list it.
The price needs no caveat. At $0.009 per 1K image on BFL's API, Klein costs under a fifth of FLUX 3 Image's $0.048, which ranks 13th on Artificial Analysis at 1117 Elo, and about a quarter of Nano Banana 2.1's $0.0336, fourth at 1176. A 4K image is $0.036. BFL reports median API times of 4.54 seconds for generation and 7.18 for edits, including prompt upsampling.
Running it locally is harder than the 8GB headline suggests. That figure is for the INT8 distilled build in ComfyUI or Diffusers, which NVIDIA timed at 8.9 seconds per 1-megapixel image on an RTX 4070. The model card's reference code peaks at 25 to 31GB on a B200 with the upsampler switched off, and the card tells you never to switch it off. Diffusers does not load the upsampler for you, though any LLM given BFL's published system prompts can stand in.
The licence is narrower than "free under $5M" sounds. The revenue test counts affiliates, so a team inside a large group fails it. The card says the exception is not blanket permission for commercial fine-tuning or distribution, and the acceptable-use terms bar surveillance, biometric processing, military use and training a competing model on the outputs. Paid self-hosting tiers are listed by volume, 10,000 images a month on Builder and 100,000 on Platform, without public prices.
That makes the older FLUX.2 [klein] 4B, under Apache 2.0, and Z-Image Turbo, 71st on Artificial Analysis at 936 Elo for $0.005 an image, the safer bets for a startup that expects to grow. On BFL's evidence Klein is the better model of the three; on its licence it is the one whose terms change as the company does.
Where it earns its place
A platform engineer at a 30-person e-commerce startup
Generate and edit 200,000 product images a month through an API
At $0.009 per 1K image that is $1,800 a month, or $2,800 at 2K, against $9,600 for FLUX 3 Image at its 1K price. Edits cost the same as generations, and bounding boxes confine an edit to one region.
A two-person indie game studio with a 12GB RTX 4070
Produce concept art in a house style on its own hardware
The INT8 distilled build in ComfyUI takes about 9 seconds an image on NVIDIA's figures, and a LoRA trained on 20 images with BFL's recipe takes about 4.5 hours on one GPU. Revenue under $5M makes commercial output free; the upsampler can run as a hosted LLM to save VRAM.
An ML lead at a retailer with more than $5M in revenue
Self-host an editing model for catalogue work
The Kommunity Licence does not cover the company, so it needs a Builder (10,000 images a month) or Platform (100,000) licence, priced on request, or the API. For a fixed monthly volume the API price is known in advance; the licence is not.
Before you commit
- Setup
- Accept the Hugging Face gate; the reference CLI loads BFL's 9B upsampler and an NSFW classifier by default. In Diffusers you load an upsampler yourself. The API model is flux-3-image-klein, polled until ready.
- Integrations
- ComfyUI, Diffusers, SGLang Diffusion, vLLM-Omni, fal and BFL's API; NVIDIA TensorRT-LLM builds; Ostris AI Toolkit for LoRA training.
- Data and privacy
- FLUX Kommunity Licence: commercial output free under US$5M revenue including affiliates; commercial fine-tuning or distribution needs a licence; no surveillance, biometric, military or competing-model training use. BFL's API retention terms were not on the docs pages read.
- Cost at scale
- 100,000 API images a month is $900 at 1K or $3,600 at 4K. Self-hosted, a 1-megapixel image takes 0.55 seconds on a GB300 or about 4 seconds on an RTX 5090, per NVIDIA's tests.
- Limits
- Up to 4K output and 10 reference images; peak VRAM 25 to 31GB on the reference code before the upsampler; serving a LoRA adds about 5% to generation time; no independent ranking yet.
What's good
- Generation, editing and 10-image references in one 5.6B model
- API from $0.009 per image, $0.036 at 4K
- Base checkpoint for LoRAs; BFL's recipe trains one in about 4.5 hours on one GPU
- Day-one support in ComfyUI, Diffusers, SGLang and vLLM-Omni
What's not
- Quality claims are vendor-run; not yet on Artificial Analysis
- Free commercial use stops at $5M revenue including affiliates
- Needs a prompt upsampler, which adds about 28GB of GPU memory
- Self-hosting licence prices not published
Best for
Small companies and developers who want a current open image model with editing built in, or a very cheap API one.
Alternatives
FLUX 3 Image Klein: common questions
- Is FLUX 3 Image Klein worth it?
- Klein is the cheapest credible route to FLUX 3-style editing and 4K output, though every claim about its quality so far is Black Forest Labs' own. We rate it 4.0 out of 5.
- How much does FLUX 3 Image Klein cost?
- Weights free under $5M annual revenue. BFL API: $0.008 (768px), $0.009 (1K), $0.011 (1.5K), $0.014 (2K), $0.036 (4K) per image, generation or editing.
- What are the downsides of FLUX 3 Image Klein?
- Quality claims are vendor-run; not yet on Artificial Analysis; free commercial use stops at $5M revenue including affiliates; needs a prompt upsampler, which adds about 28GB of GPU memory; self-hosting licence prices not published.
- Who is FLUX 3 Image Klein best for?
- Small companies and developers who want a current open image model with editing built in, or a very cheap API one.
- How much does FLUX 3 Image Klein cost for a team or in production?
- 100,000 API images a month is $900 at 1K or $3,600 at 4K. Self-hosted, a 1-megapixel image takes 0.55 seconds on a GB300 or about 4 seconds on an RTX 5090, per NVIDIA's tests.
- Is FLUX 3 Image Klein safe for company data?
- FLUX Kommunity Licence: commercial output free under US$5M revenue including affiliates; commercial fine-tuning or distribution needs a licence; no surveillance, biometric, military or competing-model training use. BFL's API retention terms were not on the docs pages read.
- What are the best FLUX 3 Image Klein alternatives?
- The closest alternatives we have reviewed are Qwen-Image-2.1, Z-Image-Turbo, FLUX 3 and Nano Banana 2.1.
Answers checked Oct 22, 2026.
Make FLUX 3 Image Klein? Show this rating on your site
