Black Forest Labs releases a 5.6 billion parameter open weights image model that runs on 8GB consumer GPUs

Black Forest Labs released FLUX 3 Image Klein, a 5.6 billion parameter open weights model for text-to-image generation, image editing, and multi-reference generation using up to 10 reference images. The model delivers native 4K generation and bounding-box controls for precise element placement, running on consumer GPUs with as little as 8GB of VRAM.

The release includes a base model for fine-tuning and a step-distilled model for low-latency generation. FP8, INT8, and NVFP4 builds arrive with NVIDIA support under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue under $5 million.

Smaller model, higher benchmarks

In internal evaluations, FLUX 3 Image Klein ranked first in text-to-image, multi-reference generation, and image editing against larger open weights models. It also rated above closed models Ideogram V4.5 and FLUX.2 [pro]. At 5.6B parameters, it is the smallest model in the comparison - Qwen-Image 2.1 runs 7B parameters and Krea 2 uses 12.9B.

Compared with FLUX.2 [klein] at 9B parameters, the new model gains 91 Elo in text-to-image and 70 in image editing, with 38% fewer parameters. The generator weights require 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 build, the model runs on NVIDIA GeForce RTX GPUs through ComfyUI or Diffusers.

Generation speed across hardware

The distilled model with INT8 build generates a 1MP image in about 4 seconds on an NVIDIA GeForce RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI. On NVIDIA GB300, the distilled model produces a 1MP image in under 0.6 seconds using NVFP4. Serving a LoRA adds roughly 5% to generation time.

Through the BFL API, FLUX 3 Image Klein delivers lower latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5, and Krea 2, while achieving the highest quality rating at its latency tier. API pricing starts at $0.009 per 1K image, with prompt upsampling and pixel restoration that preserves unedited image regions exactly.

Training infrastructure and fine-tuning support

The model was trained on NVIDIA GB200 NVL72 and optimized for NVIDIA's accelerated computing platform. Day-one support covers ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni. LoRA training works through Ostris's AI Toolkit, and NVIDIA NeMo AutoModel ships recipes that work directly from the released checkpoints. For teams working with Generative AI Courses or building custom image pipelines, the base checkpoint is built specifically for fine-tuning on proprietary datasets.

Safety approach

Black Forest Labs applied dataset filters during pre-training to limit exposure to harmful content and used targeted fine-tuning during post-training to suppress harmful behaviors. The checkpoints were red teamed internally and evaluated by external safety partner Cinder before release. The company said the model demonstrated "substantially fewer vulnerabilities than other popular open weights models."

Why this matters for creatives and marketers

FLUX 3 Image Klein runs locally on a laptop with 8GB of VRAM, which means creative teams can edit images on-device without sending assets to the cloud. That matters for data sovereignty, client confidentiality, and iteration speed. The bounding-box controls let you place elements or confine edits to a defined area - useful for compositing product shots or generating variations within brand templates. At $0.009 per 1K images through the API, the economics work for high-volume content pipelines. The model also accepts up to 10 reference images, letting you steer output toward a specific visual style without writing lengthy prompts.