Black Forest Labs' FLUX 3 Klein Runs Image Editing on 8 GB GPUs
Black Forest Labs released a 5.6B open-weights image model that handles text-to-image, editing, and multi-reference composition, running on 8GB GPUs.
- Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weights image model for generation, editing, and multi-reference composition.
- Accepts up to 10 reference images in one call, uses Qwen3-VL-4B as text encoder.
- Runs on 8GB consumer GPUs via INT8; roughly 4 seconds per 1MP image on RTX 5090.
- Bounding box layout control inside prompts on a 0-1000 grid for placement and edits.
- Free commercial use for teams under $5M revenue; API from $0.009 per 1K image.
- Ships base, distilled, FP8, INT8, NVFP4 variants; full details here.
FLUX 3 Image Klein Brings Generation and Editing to Consumer GPUs
Black Forest Labs has released FLUX 3 Image Klein, an open-weights image system that handles text-to-image generation, image editing, and generation with as many as 10 reference images. The company says quantized configurations can run on GPUs with 8 GB of VRAM.
The 5.6-billion-parameter image transformer is smaller than its predecessor by 38 percent. Its downloadable weights support local deployment and fine-tuning, subject to a custom commercial license. Black Forest Labs also offers the distilled checkpoint through its API.
Three workflows share one transformer
Most image systems use separate models for generation, inpainting, and reference-guided composition. Klein uses one image transformer for prompt-only generation, single-reference editing, and multi-reference composition.
Developers can combine up to 10 reference images for character consistency, style transfer, or scene assembly. The same image weights process each mode, reducing the model switching and orchestration required by chained pipelines.
The “5.6B” label covers the image transformer. Klein also uses Qwen3-VL-4B-Instruct as a separate text encoder. That vision-language model converts prompts and reference images into conditioning signals the image transformer can process.
Choose between control and speed
| Configuration | Precision | Sampling | Primary use |
|---|---|---|---|
| Klein Base | BF16 | 50 steps, guidance 3.5 | Fine-tuning, low-rank adapter training, and research |
| Klein distilled | BF16 | 16 steps, no guidance | Generation and editing; also used by the Black Forest Labs API |
| Quantized builds | NVFP4, FP8, or INT8 | Runtime-dependent | Lower-memory deployment and faster inference |
BF16 stores values in a 16-bit format, while FP8, INT8, and NVIDIA’s NVFP4 reduce memory use with lower-precision representations. NVIDIA helped develop the quantized builds.
Step distillation trains the production checkpoint to approximate the base model’s longer denoising process in fewer passes. Black Forest Labs reports comparable output quality with roughly three times the sampling throughput.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
