Black Forest Labs launched FLUX 3 Picture Klein on October 15, 2026, a 5.6B-parameter open-weights mannequin for text-to-image technology, picture modifying, and multi-reference technology with assist for as much as 10 reference pictures.
Capabilities and Mannequin Builds
The mannequin shares core capabilities with the total FLUX 3 Picture mannequin: native 4K technology, pixel-perfect modifying, and bounding bins that management the place components are positioned or confine edits to an outlined space. Black Forest Labs stated the mannequin builds on the FLUX 3 household’s shared learnings throughout pictures, video, and audio — pictures contribute spatial element, video provides movement and bodily dynamics, and audio gives details about occasions and interactions — and that these alerts assist the mannequin generate pictures grounded in the actual world.
The discharge features a base mannequin meant for fine-tuning and a step-distilled mannequin for low-latency technology, alongside FP8, INT8, and NVFP4 builds developed with NVIDIA assist. The generator weights require 11.2 GB of reminiscence in BF16 and 5.6 GB in FP8, and with the INT8 construct the mannequin runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM by means of ComfyUI or Diffusers. The corporate stated the lowered parameter rely makes the mannequin sensible to run regionally, cheaper to adapt, and suited to on-device picture modifying immediately on laptops.
Reported Analysis and Efficiency Outcomes
In inside evaluations, Black Forest Labs reported, FLUX 3 Picture Klein ranks first in text-to-image, multi-reference technology, and picture modifying, forward of bigger open-weights fashions, and charges above the closed fashions in its comparability, Ideogram V4.5 and FLUX.2 professional. At 5.6B parameters, it’s the smallest open mannequin in that comparability, towards 7B for Qwen-Picture 2.1 and 12.9B for Krea 2. In contrast with the corporate’s FLUX.2 klein 9B, the brand new mannequin good points 91 Elo in text-to-image and 70 in picture modifying whereas utilizing 38% fewer parameters, based on the corporate.
The corporate stated FLUX 3 Picture Klein additionally extends the pace–high quality Pareto frontier, delivering decrease latency by means of the BFL API than Qwen-Picture 2.1, FLUX.2 professional, Ideogram V4.5, and Krea 2, with a reported median API latency of 4.54 seconds for text-to-image and seven.18 seconds for modifying at 1MP. Below the acknowledged methodology, latency was measured finish to finish from request submission to the completed 1MP picture, together with queueing; FLUX 3 Picture Klein was timed on the corporate’s servers, whereas the opposite fashions have been timed from the shopper, together with polling and community time. Krea 2 was timed on its Medium tier, and its Giant tier took 31.1 seconds for text-to-image.
For native inference, the corporate reported that the distilled mannequin with the INT8 construct generates a 1MP picture in about 4 seconds on an NVIDIA GeForce RTX 5090 and in beneath 9 seconds on an RTX 4070 by means of ComfyUI. Timing charts revealed with the discharge, measured by NVIDIA with RTX outcomes recorded in ComfyUI, present the 16-step distilled mannequin starting from 0.55 seconds on a GB300 to eight.9 seconds on an RTX 4070 for 1024×1024 text-to-image, and the 50-step base mannequin starting from 1.7 seconds on GB300 and DGX Station to 43.8 seconds on DGX Spark. The charts present every GPU’s quickest construct and measure technology time on the GPU itself moderately than end-to-end API latency.
Availability, Licensing, and Security
The distilled mannequin is out there within the BFL API as flux-3-image-klein with immediate upsampling and pixel restoration, which the corporate stated preserves unedited areas of a picture precisely, and pricing begins at $0.009 per 1K picture. Base and distilled weights are posted on Hugging Face, reference inference code is out there on GitHub with a Gradio demo for bounding-box management, and a demo is out there on Hugging Face Areas.
Day-one assist spans ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni, and the mannequin was skilled on NVIDIA GB200 NVL72 and optimized for NVIDIA’s accelerated computing platform. The corporate stated that on NVIDIA GB300 the distilled mannequin generates a 1MP picture in beneath 0.6 seconds in NVFP4, and that serving a LoRA provides round 5% to technology time.
The bottom checkpoint is constructed for fine-tuning: Diffusers and Ostris’s AI Toolkit assist LoRA coaching from day one, and NVIDIA’s NeMo AutoModel ships coaching recipes that run immediately from the launched checkpoints. The weights are launched beneath the FLUX Kommunity License, which allows free business use for people and firms with annual income beneath $5M, alongside analysis and non-commercial use; bigger organizations can deploy commercially by means of self-hosted licensing choices or the API.
Black Forest Labs stated it applies dataset filters throughout pre-training to restrict publicity to dangerous content material and focused fine-tuning throughout post-training to suppress dangerous behaviors, and that the FLUX 3 Picture Klein checkpoints have been pink teamed internally and evaluated by exterior security accomplice Cinder, demonstrating considerably fewer vulnerabilities than different common open-weights fashions. FLUX 3 Picture Klein is the second open-weights launch from the FLUX 3 household, following FLUX 3 Motion, the corporate’s open-weights 7B mannequin for robotic management, and the corporate stated extra releases will comply with.
