FLUX 3 Image Klein ships open weights for 8 GB GPUs

Black Forest Labs released open weights for FLUX 3 Image Klein, a 5.6-billion-parameter image model that runs on consumer GPUs with as little as 8 GB of video memory, in a blog post. VentureBeat reported the release at 8:50 a.m. PT on October 15, 2026.

The weights sit on Hugging Face as a base checkpoint for fine-tuning and a step-distilled checkpoint for low-latency generation. The German lab also published FP8, INT8 and NVFP4 builds made with NVIDIA support, and trained the model on NVIDIA GB200 GPUs in an NVL72 system.

FLUX 3 Image Klein is the open-weights member of the FLUX 3 Image family, which Black Forest Labs announced in July 2026. Klein does text-to-image generation, single-reference editing and multi-reference composition with up to 10 reference images in one set of weights, at native 4K, with bounding boxes on a 0 to 1000 grid that set where an element lands.

Black Forest Labs measured the 16-step distilled checkpoint in its INT8 build at about 4 seconds for a one-megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070, both through ComfyUI. The generator weights take 11.2 GB in BF16 and 5.6 GB in FP8.

The terms are where the release narrows. Under the FLUX Kommunity License v.1.0, the eligibility definition is short: "Qualifying User" means you, if at any time you together with your affiliates' gross annualized revenue is less than US$5,000,000. Only Qualifying Users get the commercial-output exception, and it carries conditions, including content filtering or output review before distribution and AI disclosure where the law requires it. A business above the cap needs a commercial license to self-host, or can call the API, where Klein costs $0.008 at 768x768, $0.009 at 1K, $0.014 at 2K and $0.036 at 4K per image, with generation and editing priced the same.

Day-one support covers ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training through Ostris's AI Toolkit and recipes in NVIDIA NeMo AutoModel, per ComfyUI Wiki. In Black Forest Labs' internal evaluations, Klein gains 91 Elo in text-to-image over FLUX.2 [klein] 9B with 38 percent fewer parameters.

For a video workflow, that is a local way to turn out keyframes, reference plates and thumbnails without paying per image, as long as the project stays under the revenue cap.

The weights are on Hugging Face: update ComfyUI, accept the terms on the repository, then drop the base or distilled build into your models folder.