FLUX 3 Image Klein: frontier image model you can actually run

Black Forest Labs, the Freiburg lab founded by the researchers behind Stable Diffusion, released the open weights for FLUX 3 Image Klein on October 15, 2026. It is the lab's most capable open-weights image model to date: text-to-image generation, pixel-perfect image editing, and multi-reference generation with up to 10 reference images, all in one 5.6-billion-parameter package.

It carries the headline features of the full FLUX 3 Image model: native 4K generation and bounding boxes that control where elements are placed or confine an edit to a defined region. The release ships two ways — a base model for fine-tuning and research, plus a step-distilled model for low-latency generation — along with FP8, INT8, and NVFP4 quantized builds developed with NVIDIA.

The story here is the size. At 5.6B parameters, the INT8 build runs on NVIDIA GeForce RTX cards with as little as 8GB of VRAM through ComfyUI or Diffusers — meaning a standard gaming laptop can edit images locally instead of shipping them to an API. BFL reports the distilled INT8 model generates a 1MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070.

In BFL's internal evaluations, Klein ranks first among leading open-weights image models in text-to-image, multi-reference generation, and editing — ahead of larger models like Qwen-Image 2.1 (7B) and Krea 2 (12.9B) — while costing less through the BFL API. Against the lab's own FLUX.2 klein 9B, it gained 91 Elo in text-to-image and 70 in editing with 38% fewer parameters. Treat those numbers as company claims until independent testing lands, but the direction is clear: frontier image quality is getting small enough to own.

The weights ship under the FLUX Kommunity License: free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Larger organizations deploy commercially through BFL's self-hosted licensing or the API, where the distilled model (flux-3-image-klein) starts at $0.009 per 1,000 images.

Day-one tooling is unusually complete: ComfyUI, Hugging Face Diffusers, vLLM-Omni, SGLang Diffusion, and Ostris's AI Toolkit for fine-tuning, with NVIDIA NeMo AutoModel training recipes that run straight from the released checkpoints. Base and distilled weights are on Hugging Face, reference inference code and a Gradio demo for bounding-box control are on GitHub, and there's a live demo on Hugging Face Spaces. On safety, BFL says the checkpoints passed internal red teaming plus evaluation by external partner Cinder with fewer vulnerabilities than other popular open-weights models.

This is the second open-weights release from the FLUX 3 family, following September's FLUX 3 Action for robotics — BFL says more are coming. The lab's pitch, from co-founder and CEO Robin Rombach, is that open weights let a global community adapt one foundation to very different needs, from creative generation to robot control.

Practically: if you build anything that generates or edits images, a model this capable at 8GB of VRAM changes the calculus. You can fine-tune on your own style, keep everything local for data-sensitive work, and deploy without per-image costs. The 5.6B size is also small enough to experiment cheaply with LoRA training — BFL notes adding a LoRA costs only about 5% in generation time. Download the weights, try the distilled INT8 build in ComfyUI, and see how it handles your actual images before deciding.