All posts

Blog

Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image model

FindAmNow
Black Forest LabsFLUXimage generationopen-weight modelsgenerative AINVIDIA

On 15 October 2026 Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for image generation, editing and multi-reference work. The company says an INT8 build runs on NVIDIA RTX cards with 8 GB of video memory. Quality results are BFL's own, and free commercial use is limited to companies under $5 million in revenue.

Black Forest Labs (BFL), the German company behind the FLUX image models, released FLUX 3 Image Klein on 15 October 2026. It is a 5.6-billion-parameter open-weights model that can create images from text, edit existing images and combine up to ten reference images into one picture.

BFL says the model is small enough to run on ordinary gaming graphics cards. It is the second open-weights release in the FLUX 3 family, after FLUX 3 Action.

What the model can do

According to BFL, FLUX 3 Image Klein shares core features with the full FLUX 3 Image model:

  • Text-to-image generation, with native output up to 4K resolution.
  • Image editing that BFL describes as "pixel-perfect", leaving untouched parts of an image unchanged.
  • Multi-reference generation with up to ten input images, for example placing a person from one photo into the setting of another.
  • Bounding boxes, which let users mark where each element should appear, or limit an edit to one area of the picture.

The Hugging Face team's guide shows how the bounding boxes work in practice. Users add a short list of boxes to the prompt, each with a description, and can also tell the model to move an object from one spot in a reference image to another.

Two versions, built for small hardware

The release includes two checkpoints:

  • A distilled model for fast generation, which uses 16 steps by default.
  • A base model with 50 steps, intended for fine-tuning, meaning further training on a user's own images. LoRA training, a cheap way to teach a model a new style or subject, is supported from day one.

BFL also released compressed FP8, INT8 and NVFP4 builds made with NVIDIA's help. The company says the model's weights need 11.2 GB of memory at full BF16 precision and 5.6 GB in FP8. With the INT8 build, BFL says it runs on NVIDIA GeForce RTX cards with as little as 8 GB of video memory.

On speed, BFL reports, using NVIDIA's measurements, that the distilled model makes a one-megapixel image in about 4 seconds on an RTX 5090 and in under 9 seconds on an RTX 4070. Hugging Face measured about 2 seconds for a 1024 × 1024 image on a data-centre H100 GPU.

BFL's quality claims

BFL says FLUX 3 Image Klein ranks first among open-weights models for text-to-image, multi-reference generation and editing in its internal evaluations. It also says the model rates above two closed models in the comparison, Ideogram V4.5 and BFL's own FLUX.2 [pro]. Compared with the earlier FLUX.2 [klein] 9B, BFL reports gains of 91 Elo points in text-to-image and 70 in editing, with 38% fewer parameters.

These are BFL's own results from its own tests. BFL notes that its parameter counts cover only the image generator, not the text encoder, which also has to run.

License and safety

The weights are released under the FLUX Kommunity License. It allows free commercial use for individuals and companies with annual revenue under $5 million, as well as research and non-commercial use. Larger companies need a paid self-hosted license or must use BFL's API.

BFL says it filtered harmful content out of the training data and added safety training after pre-training. It says the model was red-teamed internally and tested by an outside safety partner, Cinder, and that it showed "substantially fewer vulnerabilities" than other popular open models. That comparison is also BFL's own.

Why it matters

Image models that fit on a consumer graphics card let designers, small studios and developers generate and edit images on their own machines, without sending files to a cloud service. BFL also points to companies with data-sovereignty rules, which may need to keep work on their own systems.

The revenue cap in the license is the main limit: larger businesses cannot use the free weights commercially.

Sources

Source