Black Forest Labs' 5.6B open-weights image model — runs locally on an 8GB GPU, with native 4K, pixel-perfect editing, up to 10 reference images, and bounding-box layout.
Last verified · 2026-10-16 · by Moe Ameen
FLUX 3 Image Klein is Black Forest Labs' open-weights image model, released October 15, 2026. It is a 5.6B-parameter model (the count covers the image generator, not the text encoders) for text-to-image generation, image editing, and multi-reference generation — the open, run-it-yourself counterpart to the full FLUX 3 Image, and the second open-weights release in the FLUX 3 family after FLUX 3 Action. Black Forest Labs is the German studio founded by several of Stable Diffusion's original researchers.
The defining trait is that it runs locally on hardware you already own. Klein ships as a base model built for fine-tuning and a step-distilled model built for fast inference, with FP8, INT8, and NVFP4 builds made with NVIDIA. The generator weights are about 11.2 GB in BF16 and 5.6 GB in FP8, and the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. On an RTX 5090 the distilled INT8 model renders a 1-megapixel image in roughly 4 seconds, and under 9 seconds on an RTX 4070.
It keeps the full FLUX 3 Image feature set: native 4K generation, pixel-perfect editing, multi-reference composition with up to 10 reference images, and a bounding-box system to place named elements or confine an edit to a defined region. Weights and reference code are on Hugging Face and GitHub, with day-one support in ComfyUI and Hugging Face Diffusers; the distilled model is also on the BFL API as flux-3-image-klein, starting at $0.009 per 1,000 images. LoRA fine-tuning is supported through Diffusers, Ostris's AI Toolkit, and NVIDIA NeMo AutoModel.
The license is the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use; larger organizations deploy through self-hosted licensing or the API. Black Forest Labs reports first place among open-weights models in its own text-to-image, multi-reference, and editing evaluations, but published no independent benchmarks at release — treat the quality claims as promising but not yet third-party verified, and confirm specs, VRAM figures, and pricing on the lab's pages.
Klein changes the generation step in one specific way: you can iterate a scene locally and privately as many times as you want, editing a single region or swapping references, without a per-image meter running. That iterate-until-right loop is where Klein earns its keep — and the moment you have the frame you want, the work shifts from rendering to distributing, which is exactly where Kompozy picks up. The most useful handoff is turning that one hard-won frame into a content unit a model never assembles on its own: bring the final Klein still into Kompozy as a source and HyperFrames renders it into a brand-exact multi-slide Carousel with your copy, the same still becomes a Photo Post and a Quote Graphic, and the Persona Brief writes platform-native captions in your voice across all of it.
From there Kompozy fans the same concept into the formats a still model never produces — a HeyGen-driven Persona Short or a Listicle Video, plus a Blog Article and an Email Newsletter — all scheduled and published across eight social platforms plus blog and email from one queue. If your brand needs a recurring face, Kompozy's Gemini face-locked Persona Photos and HeyGen persona video complement Klein's scene control rather than competing with it: Klein owns the pixels inside the frame, Kompozy owns the formats, captions, voice, and the publish around it.
FLUX 3 Image Klein is Black Forest Labs' open-weights image model, released October 15, 2026. It is a 5.6B-parameter model for text-to-image generation, image editing, and multi-reference generation that runs locally on a consumer GPU, keeping the full FLUX 3 Image features: native 4K, pixel-perfect editing, up to 10 reference images, and bounding-box layout control.
Black Forest Labs says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. The weights are about 11.2 GB in BF16 and 5.6 GB in FP8. The distilled INT8 model renders a 1-megapixel image in roughly 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070.
The weights are open under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use. Larger organizations need self-hosted licensing or the BFL API, where the distilled model starts at $0.009 per 1,000 images. Confirm the current terms on Black Forest Labs' pages.
Klein renders the image but doesn't publish it. Bring the still into Kompozy to build a brand-exact carousel and quote graphic from it via HyperFrames, spin the same concept into a Persona Short or Listicle Video, write captions in your voice, and schedule and publish across eight social platforms plus blog and email from one queue.