FLUX 3 Image Klein An open-weights 5.6B-parameter image generation model that delivers frontier-model quality while running on a single consumer GPU. It supports native 4K generation, bounding-box control, and pixel-perfect editing, and was trained for high throughput on NVIDIA hardware.
Visit productFeatured media
Loading media…
View on XCosigns Received
“FLUX 3 Image Klein from @bfl_ai can: - Run on a single consumer GPU with 5.6B params - Generate at native 4K - Place things exactly with bounding boxes - Do pixel-perfect edits - Edit with multiple references - Make an image in ~4s on an RTX 5090 (INT8) Tested it myself, here's the result 👇 Prompt: A cinematic high-fashion photograph treated with the texture and lighting of a 19th-century oil portrait, featuring a young East Asian woman standing in a dark, ornate penthouse at night. She wears a structured black silk evening gown with a high collar and long sleeves, the fabric catching light along its sharp folds and open back. One hand rests on a tall window frame as she looks outward over a glittering city. Cool moonlight from the side mixes with warm interior light, creating strong chiaroscuro across her face, neck, and the gown’s silhouette. Shallow depth of field, visible film grain, rich color grading with deep shadows and selective gold highlights. Intimate yet distant mood, refined and quietly captivating. aspect_ratio:16:9 resolution:2k”
“Ostris AI Toolkit has day 0 support for FLUX 3 Image Klein. Thank you @bfl_ai for the early access and the amazing model! https://github.com/ostris/ai-toolkit/pull/1078”
“今天开源模型扎堆发布，挺热闹的。不过最令我激动的还是 Flux3，开源图片模型真的很久没有好用的模型了 激活参数 56 亿，三个版本，原版 bf16、FP8、INT8 和 NVFP4 - bf16 原版加载后占 18.8 GB，出图时最高 20.7 GB。 - FP8加载后 14.4 GB，出图时 16.3 GB。 - NVFP4加载后 13.4 GB，出图时 15.3 GB。 - INT8 目前没查到 现在很多人依旧跟我说开源模型不如闭源，一部分我同意，但我想没有开源模型的推动，也不会有类似 GPT-image 和 nanobanana 这些闭源图片模型的出现，而且开源图片模型的应用领域更广，这是我的一点个人见解，如果有不对的地方还请指正”
“FLUX 3 Image Klein、試してみました 16GBのGPUでも、実写の縦長1枚が7秒で出ます その1枚を、MiniMax H3で15秒動かしてみました 猫が指のにおいをかぐところがかわいいですね！”
“Flux 3 Image Klein is out!✨ As a creative partner of the @bfl_ai community with early access, I’ve been testing FLUX 3 Image Klein — the bounding boxes are exactly the control we’ve needed. Place elements precisely, then edit just that region while the rest of the image stays untouched. Open, fast, and surprisingly capable on consumer hardware.”
“Consumer GPU + native 4K is the part that actually changes things. Samples of the output here:”
“FLUX 3 Image Klein is out now, it's only 5.6B and we also worked on almost all quantizations - you can run it on almost all RTX cards now. Go play with it, it's also the best at Editing. We have something new coming next week! 👀”
“FLUX 3 Image Klein is here! Bounding-Box control, pixel-perfect editing, multi-ref and a lot more - with only 5.6B parameters. Weights are available on HF! This model is a massive leap forward for applications that need to run locally. Go use it 🌲”
“releasing the base checkpoint and not just the distilled one is a real gift to people who want to fine-tune on their own style”
“we just released the best open weight image model: FLUX 3 Image Klein is super efficient, richtig klein & mega nice. pls try it out and share feedback 🙏 It literally runs on everything and is ideally suited for local development.”
“Open weights at frontier quality is the combination I have been waiting for. I run image models on my own rig, and the customizable part is what sells it for me, since stock checkpoints always need my own fine-tuning to behave.”
“Already tried and the results are great 😍 It’s also kuch faster on Mac Studio then Qwen Image 2.1 Turbo Thank you!”
“native 4K and bounding boxes on a single consumer GPU is huge for small studios like ours. open weights at this quality changes the economics”
“Decent editing capabilities for a relatively small image model. Faces are now very consistent and actually better than Flux 2 Klein 9B in my tests (they match the previous 32B Dev model). I use an Ollama node and Qwen 3.5 9B for prompt enhancement. Less than a minute on a 3060.”