Flux 2 Klein: BFL's small open image models, eight months after launch

Written from Black Forest Labs' own posts and VentureBeat's coverage, plus the model cards on Hugging Face, with scores pulled from arena.ai and Artificial Analysis on September 30, 2026. Originals are linked throughout. Rankings in this corner move every few weeks, so check them before quoting anything from here.

Under half a second. That's the number that made me stop scrolling in January. Black Forest Labs put it in the [FLUX.2 [klein] announcement](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence) on January 15, 2026, and VentureBeat's story the next day put "less than a second" right in the headline. A model you can download, running on a gaming card, handing back a picture before you've let go of the mouse. A couple of years ago I'd have filed that under wishful thinking.

Part of my job at Flami is deciding which image models sit in front of people making product visuals, and most of what they generate is drafts that never get printed. So small, fast Flux 2 Klein models get more of my attention than their size suggests. Eight months on, here's where Klein actually ended up.

Promised in November, with a different license

When BFL launched the FLUX.2 family on November 25, 2025, Klein sat at the bottom of the lineup post as "coming soon," a size-distilled model under Apache 2.0. I'll admit I didn't believe it would ship without something being cut. Something was. Just not the part I was worried about.

What actually shipped in January

Four checkpoints rather than two. You get a 4B and a 9B, each as a distilled version tuned for speed plus a slower base version meant for fine-tuning. Every one of them does text-to-image and image editing in the same model, with several reference images allowed, which the older small Flux releases never managed.

Now the license. Only the 4B models are under Apache 2.0. The 9B pair ships under the FLUX Non-Commercial License, and the 4B model card spells out that split in its release notes. If you want Klein inside a paid product without a separate deal with BFL, the 4B is the one you're allowed to use.

For scale, the open [dev] model has 32 billion parameters and pairs with a 24B Mistral-3 vision-language model, according to Black Forest Labs' own FLUX.2 launch post. Klein 4B is eight times smaller. BFL says it fits in about 13 GB of VRAM, so an RTX 3090 or 4070 will do. Quantized FP8 and NVFP4 builds cut memory further. Bigger November siblings like Pro get their own coverage in our earlier look at the FLUX.2 launch.

Numbers, some of them independent

Speed first, since that's the headline. BFL quotes under 0.5 seconds on modern hardware for the distilled versions. VentureBeat mentions sub-second generation on Nvidia's GB200. Those are best-case figures, and I'd probably expect resolution and step count to push them around on real jobs.

Quality is where January and September look different. At launch all we had was BFL's word. BFL's post says Klein matches or beats models five times its size and outperforms Z-Image. Now there are blind votes to check that against.

• Artificial Analysis, text-to-image: Klein 9B at 941 over 11,520 appearances, Z-Image Turbo at 941, Klein 4B at 863, with FLUX.2 [dev] as the board's 1000 reference point;
• arena.ai, September 24 update: flux-2-klein-9b 60th with 1070 on 150,780 votes, flux-2-klein-4b 68th with 1030, z-image-turbo 57th with 1084, flux-2-dev 36th with 1146.

So the Z-Image claim doesn't survive contact with voters. It's a tie on one board and a small loss on the other, and that's the 9B. And the Apache-licensed 4B trails both.

How far is Klein from its parent? In November VentureBeat reported that FLUX.2 [dev] won 66.6% of text-to-image comparisons, against 51.3% for Qwen-Image. Klein is distilled from the same base. On Artificial Analysis the 9B sits about 60 points below dev, which I'd read as "noticeably weaker, still the same family look."

Price is the other half. That same November piece put the cloud [pro] model at $0.03 per megapixel. On the Artificial Analysis leaderboard, Klein 9B shows up at about $15 per 1,000 images through API providers, pro at $30 and Z-Image Turbo at $5. Cheap across the board. Your own GPU beats all of them once the volume is big enough, though you pay for setup and someone's time.

Who needs it when the cloud is this cheap?

Teams pushing thousands of images a day, for one. Then there's anyone who can't send product photos to someone else's server, like a brand with an unreleased launch or an agency under a strict client contract.

Klein isn't alone on that turf. Z-Image Turbo is also small and quick under an Apache 2.0 license; a separate write-up on Z-Image covers its trade-offs in more detail. BFL obviously built Klein to compete with it. For commercial work the fair matchup is Klein 4B against Z-Image Turbo, since those are the two you can ship freely, and right now Z-Image is ahead on both boards while costing less per image.

Closed models have moved as well. Those November win rates were measured against the first Qwen-Image. Its second generation is another story, with Qwen Image 2.0 Pro scoring 1030 on Artificial Analysis, above FLUX.2 [pro] at 1004. If that's the model you're weighing, there's a Qwen Image 2 page and a longer review.

Where the shrinking shows

Back in the spring I'd have bet that small text is the first thing distillation breaks: packaging copy, ingredient panels, any dense block of type. I haven't run that comparison on our own product scenes yet, so call it a prediction, not a result.

That gap might come mostly from the distillation, or simply from where any 4B model naturally lands below its own full-size version, and either way it's a fair trade for something this fast. For moodboards and layout sketches it's plenty. For the final hero shot of a product with a real label, I still reach for the full Flux 2, which is the version we run in Flami for photoreal product work. Klein is built for exactly that kind of quick, throwaway sketch, the kind nobody should have to pay for.

Klein, in short

What is Flux 2 Klein?

A family of small image models from Black Forest Labs, released January 15, 2026, distilled from the FLUX.2 base. There are 4B and 9B sizes, each in a distilled and a base version. They handle generation and editing in one model.

Can I use Flux 2 Klein commercially?

The 4B versions, yes, under Apache 2.0. The 9B versions are under the FLUX Non-Commercial License, so paid products need a separate agreement with BFL.

What GPU does Klein 4B need?

About 13 GB of VRAM, per BFL, which means an RTX 3090 or 4070 and up. Quantized FP8 and NVFP4 builds need less.

How fast is it?

BFL quotes under 0.5 seconds per image on modern hardware for the distilled models. Base versions are slower.

How does it rank against Z-Image Turbo?

On arena.ai (September 24, 2026), Z-Image Turbo scores 1084. Klein 9B has 1070, and Klein 4B has 1030. Artificial Analysis has Klein 9B and Z-Image Turbo tied at 941, with Klein 4B at 863.

Figures checked on September 30, 2026.

Sources

1. [Black Forest Labs, FLUX.2 [klein]: Towards Interactive Visual Intelligence](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence)
2. Black Forest Labs, FLUX.2: Frontier Visual Intelligence
3. [VentureBeat, Black Forest Labs launches open source Flux.2 [klein]](https://venturebeat.com/technology/black-forest-labs-launches-open-source-flux-2-klein-to-generate-ai-images-in)
4. VentureBeat, Black Forest Labs launches Flux.2 AI image models
5. Hugging Face, FLUX.2-klein-4B model card
6. Hugging Face, FLUX.2-dev model card
7. Artificial Analysis, Text to Image Leaderboard
8. arena.ai, Text-to-Image Leaderboard