FLUX 3 Image Klein release: price, VRAM, license
Black Forest Labs released FLUX 3 Image Klein on 15 October 2026: a 5.6B open-weights model for images and edits, up to 4K, at $0.009 per 1K image.
Black Forest Labs released FLUX 3 Image Klein on 15 October 2026. It is a 5.6B-parameter open-weights model that makes images from text, edits them and combines up to 10 reference images, with the 4K output and bounding boxes of the full FLUX 3 Image. BFL's API charges $0.009 for a 1K image. On Versely it costs 1 credit.
Key facts
| Item | Detail |
|---|---|
| Released | 15 October 2026 |
| Size | 5.6B-parameter image transformer, plus a separate Qwen3-VL-4B-Instruct text encoder |
| Modes | Text to image, editing, up to 10 references, bounding boxes |
| Output | 768 × 768 up to 4K |
| Variants | Distilled (16 steps) for speed; Base (50 steps) for fine-tuning |
| BFL API price | $0.008 at 768 × 768, $0.009 at 1K, $0.014 at 2K, $0.036 at 4K |
| License | FLUX Kommunity License |
| Where | BFL API, Hugging Face, ComfyUI, Diffusers, fal |
An edit costs the same as a text-to-image job on BFL's API, and 1.5K is $0.011. The quality comparisons below come from BFL's internal evaluations, not independent tests.
Sources, all read on 17 October 2026: BFL's launch post, model page, Klein docs and release notes; the Hugging Face model card; fal's text-to-image and edit pages.
What is new compared with FLUX.2 [klein]
FLUX.2 [klein] came in two sizes, 4B and 9B, published on Hugging Face in January 2026. Klein is now a FLUX 3 model, and BFL's numbers for the change are these:
- Better scores with fewer parameters. Against FLUX.2 [klein] 9B, BFL reports a gain of 91 Elo in text-to-image and 70 in image editing, "with 38% fewer parameters."
- The FLUX 3 Image feature set. BFL says Klein "shares core capabilities with the full FLUX 3 Image model: native 4K generation, pixel-perfect editing, and bounding boxes." Through BFL's API, pixel restoration "preserves unedited regions of an image exactly."
- The same request as FLUX 3 Image. The release notes say prompt, images, aspect ratio, resolution and the box format are unchanged: "A FLUX 3 Image request runs on Klein by switching the endpoint."
- Trained across media. BFL says the FLUX 3 family learns from images, video and audio: "video adds motion and physical dynamics, and audio provides information about events and interactions."
- A new license. FLUX.2 [klein] 4B is Apache 2.0. FLUX 3 Image Klein uses the FLUX Kommunity License, which allows "free commercial use for individuals and companies with annual revenue under $5M." Larger organisations need BFL's API or a self-hosted license.
In BFL's internal evaluations, Klein ranks first among open-weights models for text-to-image, multi-reference generation and editing, and also rates above two closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro]. It is the smallest model there: 5.6B, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2.
On 17 October 2026, Artificial Analysis did not list FLUX 3 Image Klein on its text-to-image or image editing leaderboards. FLUX.2 [klein] 4B and 9B are on both, so an independent comparison should follow once the new model is added.
Running it on your own GPU
This is the part of the launch aimed at people who self-host.
- Weight size. BFL says the generator weights "require 11.2 GB in BF16 and 5.6 GB in FP8."
- Smallest GPU. With the INT8 build, BFL says Klein runs on GeForce RTX cards "with as little as 8GB of VRAM, through ComfyUI or Diffusers." Its docs say the model runs on consumer GPUs "down to an RTX 3090."
- Speed. With the distilled model and INT8, a 1-megapixel image takes "about 4 seconds on an NVIDIA GeForce RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI." NVIDIA measured these figures.
- Fine-tuning. Use the Base checkpoint. Ostris's AI Toolkit and Diffusers support LoRA training from day one, and a LoRA trained on Base also loads into the distilled model.
There is one catch. Klein is trained on long scene captions, and BFL's docs say it should always run with its prompt upsampler, because "a prompt sent as typed gives noticeably weaker results." BFL's upsampler is a separate 18.8 GB download in the reference code. The model card says you can use "any LLM you prefer" with BFL's system prompt instead, and that Diffusers runs neither on its own. Budget for that step before you plan on an 8 GB card. For scale, the model card's own peak-memory figure for the reference code at 1024 × 1024 on a B200 is 31 GB in BF16 with the upsampler off.
Who it is for, and who should wait
Good fit:
- High-volume drafts, thumbnails and concept rounds, where cost per image matters more than the last bit of polish.
- Teams that must keep images on their own hardware. BFL pitches Klein at "teams with data sovereignty or local deployment requirements."
- People who train LoRAs and want a FLUX 3 base small enough to adapt cheaply.
- Developers already on FLUX 3 Image who want a cheaper endpoint for the same request.
Wait, or test first, if:
- Your company earns over $5M a year and wants to self-host. That needs a commercial license from BFL.
- You need a permissive license. FLUX.2 [klein] 4B stays Apache 2.0.
- Your GPU has no room for the prompt upsampler and you have no other LLM to write captions.
- Your images carry a lot of text. The model card warns that "text rendered may be inaccurate or subject to distortion."
- You want an independent ranking before you switch. Artificial Analysis hasn't scored it yet.
How to use it on Versely
Versely added both modes on 16 October 2026, served through fal:
| Mode | Model on Versely | Credits |
|---|---|---|
| Text to image | FLUX 3 Image Klein | 1 per image |
| Edit, up to 4 images | FLUX 3 Image Klein Edit | 1 per edit |
Both output about 1 megapixel in five aspect ratios: 1:1, 9:16, 16:9, 3:4 and 4:3. Versely doesn't offer Klein's 2K or 4K sizes or its bounding boxes. Versely leaves fal's prompt expansion switched on (fal's default), so a short prompt is rewritten into a full caption before generation, the step BFL says the model needs. In Versely's own checks on 16 October, one text-to-image job and one edit each took 11 seconds end to end. Those are single runs, not a benchmark.
- Open FLUX 3 Image Klein on the web or in the iOS app, and choose text to image or the Edit variant.
- Describe the whole scene: subject, setting, light and camera. Put any words that must appear in the image inside quotes.
- For an edit, upload the photo you want changed, plus up to three more references. Refer to them by order, "the first image" and "the second image", as fal's edit endpoint asks.
- Keep the best result. Then render it again on a larger model, or animate it with an image-to-video model.
Other Versely rows for the same jobs:
| Job | Model | Credits |
|---|---|---|
| The full FLUX 3 model, 14 aspect ratios | FLUX 3 Image | 4 |
| Edits on the full FLUX 3 model | FLUX 3 Image Edit | 4 |
| The previous Klein generation | Flux 2 Klein 9B Base | 1 |
For what the full model adds, read the FLUX 3 Image launch. For another fast editing model that went live the same day, read the Krea 3 launch.
FAQ
When was FLUX 3 Image Klein released?
Black Forest Labs released it on 15 October 2026, as open weights on Hugging Face and through its API. Versely added it on 16 October 2026. BFL calls it the second open-weights release from the FLUX 3 family, after FLUX 3 Action.
How much does FLUX 3 Image Klein cost?
On BFL's API, as read on 17 October 2026, it costs $0.008 at 768 × 768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K and $0.036 at 4K. Text to image and editing cost the same. On Versely, both modes cost 1 credit at about 1 megapixel.
How much VRAM does FLUX 3 Image Klein need?
BFL says the INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM through ComfyUI or Diffusers. The weights are 11.2 GB in BF16 and 5.6 GB in FP8. You also need the prompt upsampler, or another LLM to write the full captions the model expects.
Can I use FLUX 3 Image Klein commercially?
Under the FLUX Kommunity License, individuals and companies with annual revenue under $5M can use it commercially for free. Larger organisations deploy it through BFL's API or a self-hosted license. The model card notes that the exception is "not blanket permission for commercial fine-tuning or distribution."
Is FLUX 3 Image Klein better than FLUX.2 [klein]?
By BFL's own figures, yes: 91 Elo higher than FLUX.2 [klein] 9B in text-to-image and 70 higher in editing, with 38% fewer parameters. No independent ranking existed on 17 October 2026. FLUX.2 [klein] 4B still has the more permissive Apache 2.0 license.