Flux 2 Klein: 4B vs 9B, and When Each One Wins
FLUX.2 [klein] comes in two sizes and two training styles. Here's when the 4B beats the 9B, when Base beats distilled, the settings Black Forest Labs recommends, and eight prompts to paste in.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

Most image models make you wait. You write the prompt, you watch the spinner, and by the time the picture lands you've half forgotten what you were trying to fix. Flux 2 Klein was built to kill that wait. The catch is that "Klein" isn't one model. It's a small family with two sizes and two training styles, and picking the wrong one costs you either speed or quality you didn't need to give up.
FLUX.2 [klein] is Black Forest Labs' compact line: a 4B model and a 9B model, each shipped as a fast distilled version and an undistilled Base version. On each::labs you can run the 4B Edit endpoint plus Base endpoints at both sizes. This post is the short answer to which one you should open first, with the settings BFL itself recommends and eight prompts you can paste straight in.
What Flux 2 Klein Actually Is
Black Forest Labs released FLUX.2 [klein] on January 15, 2026, and called it its fastest image model to date. The pitch in the official announcement is blunt: generate or edit images in under half a second on modern hardware. That number comes from BFL's own benchmark on datacenter GPUs, so treat it as the ceiling of what the architecture can do, not a promise about every request you send.
The more interesting design choice is that Klein doesn't split generation and editing into separate models. One compact architecture handles text-to-image, single-reference editing and multi-reference work. You give it a prompt and nothing else, you get a new image. Give it a prompt plus up to four reference images, and it edits or composes from them. Same weights, different job.
Underneath, both sizes are rectified flow transformers. The 9B pairs its 9 billion parameter backbone with an 8B Qwen3 text embedder, a heavy reader for a compact model. BFL claims the 9B matches or exceeds models five times its size. The 4B is the lightweight one, and BFL says it fits in about 13GB of VRAM, which puts it on RTX 3090 or 4070 class cards if you ever run it locally.

Flux 2 Klein 4B vs Flux 2 Klein 9B: Size Is a Quality Dial
Think of the two sizes as a dial, not a ladder. BFL's own docs frame the 4B as the model for high volume and local deployment, and the 9B as the balanced pick for quality and speed. Both distilled versions run at four inference steps. So the difference isn't how many passes the model takes. It's how much model is doing each pass.
In practice that shows up in three places. Prompt reading is the first: the 9B brings that 8B Qwen3 embedder to long, layered descriptions with several subjects and specific spatial relationships. Composition is the second, and BFL positions the 9B as the stronger choice for complex scenes. Footprint is the third. The 9B card lists about 29GB of VRAM and RTX 4090 class hardware, more than double the 4B.
Licensing is also split by size, and you should know it before you build anything around the open weights. BFL publishes the 4B and 4B Base under Apache 2.0. The 9B and 9B Base use the FLUX Non-Commercial License. If your plans involve downloading weights and shipping them, that difference decides more than any benchmark will.
My take: the 4B is underrated. For product edits, background swaps and recolors, where the image already carries most of the information, a small model with a good reference gets you most of the way there. The 9B earns its weight when the prompt is doing the heavy lifting.

Base vs Distilled: The Choice That Matters More
Here's the part most comparisons skip. Size matters less than training style. The distilled models (plain "4B" and "9B") were trained with step and guidance distillation so they finish in four steps. The Base models were trained without either, which BFL describes as keeping the full training signal.
That gives Base two traits. It's slower, because it needs many more steps to converge. And it's more varied. BFL rates Base 4B's output diversity as high and Base 9B's as the highest in the family. Run the same prompt four times on a distilled model and you get close cousins. Run it on Base and you get four real directions.
BFL pitches Base for fine-tuning, LoRA training, research and custom pipelines, the cases where, in its words, control matters more than speed. For everyday creative work, read that as: pick Base when you're exploring, and distilled when you already know what you want and need it fast.
Which Klein Variants Run on each::labs
Five Klein endpoints are live today, and they map neatly onto the two decisions above.
For fast editing, the Klein 4B Edit endpoint is the distilled 4B. It takes one to four source images, defaults to four steps, and has an acceleration setting if you want to push speed further. This is the one to wire into anything interactive.
For the 4B Base, you get both modes: Klein 4B Base Text to Image and Klein 4B Base Edit. Both expose an inference steps control from 1 to 50.
For the biggest Klein, there's Klein 9B Base Text to Image and Klein 9B Base Edit. These give you the fullest set of controls: steps, guidance scale, a negative prompt field and the same acceleration option. If you're comparing Klein against other generators, the text-to-image catalog puts them side by side behind the same API.

Settings That Come From BFL, Not Guesswork
Every number in this section comes from Black Forest Labs' official model cards, so you're starting from the maker's defaults rather than forum folklore.
Distilled models want few steps and almost no guidance. BFL's examples for both the 4B and 9B distilled cards use 4 inference steps with a guidance scale of 1.0 at 1024 by 1024. Don't crank the steps on the 4B Edit endpoint expecting a better picture. The model was distilled to finish in four. Adding passes mostly adds time.
Base models want the opposite. BFL's examples for Base 4B and Base 9B use 50 inference steps with a guidance scale of 4.0. On each::labs the 9B Base endpoints start at 28 steps and a guidance of 5, which is a reasonable middle. If a Base result looks soft or unfinished, move steps toward BFL's 50 before you touch anything else. The 4B Base endpoints start at 4 steps, so raise that number when you want true Base behavior from them.
Two more rules from BFL's docs. Klein has no prompt upsampling, so it won't quietly rewrite a lazy prompt into a good one. Write the detail yourself. And when you need a specific color, use hex codes tied to an object, like "the mug in color #FF0000", rather than a loose list of codes floating at the end. Lock the seed when you're comparing settings, otherwise you're comparing luck.
Flux 2 Klein Prompts for Text-to-Image
Because there's no upsampling, Klein rewards prompts that read like a photographer's brief: subject, setting, light, lens, mood. Here are four built that way. They run on either Base text-to-image endpoint; the dense ones are where the 9B pulls ahead.
A ceramic pour-over coffee set on a walnut counter at 7am, soft window light from the left, steam curling from the carafe, shallow depth of field, shot on a 50mm lens, muted earth tones with the kettle in color #1B6B6F, editorial product photography.
An elderly bookbinder in a cluttered workshop stitching a leather spine, warm tungsten lamp overhead, dust visible in the light beam, shelves of half-finished books behind her, documentary style, natural skin texture, 35mm film grain.
A rain-soaked night market in a narrow alley, paper lanterns overhead, a vendor flipping skewers over open flame, reflections in puddles, three customers waiting in line with umbrellas, cinematic wide shot, teal and amber color palette.
A minimalist poster layout: a single orange in color #F28C28 centered on an off-white background in color #F4F1EA, hard shadow falling to the right, large empty space at the top for a headline, flat studio lighting, clean graphic composition.
Notice what's absent: no "masterpiece", no "8k, ultra detailed". Klein reads descriptions, not incantations. Spend those words on light and position instead.

Editing and Multi-Reference Prompts
Editing is where Klein feels fastest, and where the 4B Edit endpoint makes the strongest case for itself. BFL's editing guidance boils down to a simple rule: say what changes, say what stays, and when there are several inputs, say what role each image plays. Klein accepts up to four references.
Replace the background with a sunlit concrete loft with large industrial windows. Keep the sneaker, its laces, logo placement and shadow exactly as they are. Match the new lighting direction to the existing shadow.
Change the sofa fabric to a deep green velvet in color #1F4D3A. Keep the room layout, the rug, the plants and the window light unchanged.
Use the woman from image 1, dress her in the trench coat from image 2, and place her on the street from image 3. Keep her face, hair and pose from image 1. Overcast daylight, full-body shot at eye level.
Restyle this photo as a hand-printed risograph poster with two inks, blue and fluorescent pink, visible grain and slight misregistration. Keep the composition and the subject's silhouette the same.
For one-off fixes, start on the distilled 4B Edit and only move up if the result loses detail you care about. For a multi-reference composite with a person whose identity has to hold, the 9B Base Edit endpoint with more steps is the safer bet. A good habit for either: one change per request when precision matters. Stack five edits into one prompt and you won't know which one broke.
When to Pick Which Klein
Here's the verdict, by job.
Pick the distilled 4B when speed is the product. Live previews, in-app editors, batch recolors, background swaps on a catalog: anywhere a person or a pipeline is waiting on the answer. It's also the Apache 2.0 option if you later want the open weights in your own stack.
Pick 4B Base when you want variety without the heaviest model. It's a good exploration tool for moodboards and early concepts, and BFL positions it for LoRA training and custom pipelines.
Pick 9B Base when the prompt is long, the scene is crowded, or the picture is the final asset. It has the most parameters in the family, its diversity is the highest in the family, and the extra guidance and negative prompt controls give you more to steer with.
A workflow that uses all three works better than loyalty to one. Explore on Base, settle on a direction, then hand the polish edits to the 4B Edit endpoint where each round is quick. If you chain those steps often, build it once as an each::labs flow.
When You've Outgrown Klein
Klein is built for speed and size. Some jobs need neither. When the brief is a hero image, dense typography or a scene with a dozen precise elements, step up to FLUX 3 Image, Black Forest Labs' larger generation. Our FLUX 3 Image explainer covers what changed. And if your work is mostly retouching, the roundup of the best AI image editing models shows where Klein sits next to the heavier editors.
The point isn't to graduate from Klein. It's to stop asking one model to do everything. Fast drafts on Klein, finals on something bigger, all on the same API.
Frequently Asked Questions
Is flux 2 klein 4b or flux 2 klein 9b better?
Better depends on what you're waiting for. The 9B reads complex prompts more faithfully and produces stronger compositions, which BFL frames as the balanced quality pick. The 4B is lighter, built for high volume and local deployment, and handles most single-image edits well. Start on 4B for edits and 9B for dense text-to-image.
How many steps should I use with Flux 2 Klein?
Four for the distilled models, around 50 for Base. Those are the values in BFL's official model cards, with a guidance scale of 1.0 for distilled and 4.0 for Base. On each::labs, the 9B Base endpoints default to 28 steps, so raise them if results look unfinished.
Can FLUX.2 [klein] edit images with multiple references?
Yes, up to four. BFL built Klein to handle text-to-image, single-reference editing and multi-reference composition in one model. Name each image's role in the prompt ("the coat from image 2") so the model knows what to take from where.
What license do the Klein models use?
It splits by size. BFL releases the 4B and 4B Base open weights under Apache 2.0, and the 9B and 9B Base under the FLUX Non-Commercial License. Check BFL's license text before you deploy the weights yourself.



