Black Forest Labs 發佈 5.6B 參數開放權重圖像模型 FLUX 3 Image Klein

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, a 5.6B-parameter open-weights model for text-to-image generation, image editing, and multi-reference generation with support for up to 10 reference images.

Capabilities and Model Builds

The model shares core capabilities with the full FLUX 3 Image model: native 4K generation, pixel-perfect editing, and bounding boxes that control where elements are placed or confine edits to a defined area. Black Forest Labs said the model builds on the FLUX 3 family’s shared learnings across images, video, and audio — images contribute spatial detail, video adds motion and physical dynamics, and audio provides information about events and interactions — and that these signals help the model generate images grounded in the real world.

The release includes a base model intended for fine-tuning and a step-distilled model for low-latency generation, alongside FP8, INT8, and NVFP4 builds developed with NVIDIA support. The generator weights require 11.2 GB of memory in BF16 and 5.6 GB in FP8, and with the INT8 build the model runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. The company said the reduced parameter count makes the model practical to run locally, cheaper to adapt, and suited to on-device image editing directly on laptops.

Reported Evaluation and Performance Results

In internal evaluations, Black Forest Labs reported, FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation, and image editing, ahead of larger open-weights models, and rates above the closed models in its comparison, Ideogram V4.5 and FLUX.2 pro. At 5.6B parameters, it is the smallest open model in that comparison, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. Compared with the company’s FLUX.2 klein 9B, the new model gains 91 Elo in text-to-image and 70 in image editing while using 38% fewer parameters, according to the company.

The company said FLUX 3 Image Klein also extends the speed–quality Pareto frontier, delivering lower latency through the BFL API than Qwen-Image 2.1, FLUX.2 pro, Ideogram V4.5, and Krea 2, with a reported median API latency of 4.54 seconds for text-to-image and 7.18 seconds for editing at 1MP. Under the stated methodology, latency was measured end to end from request submission to the finished 1MP image, including queueing; FLUX 3 Image Klein was timed on the company’s servers, while the other models were timed from the client, including polling and network time. Krea 2 was timed on its Medium tier, and its Large tier took 31.1 seconds for text-to-image.

For local inference, the company reported that the distilled model with the INT8 build generates a 1MP image in about 4 seconds on an NVIDIA GeForce RTX 5090 and in under 9 seconds on an RTX 4070 through ComfyUI. Timing charts published with the release, measured by NVIDIA with RTX results recorded in ComfyUI, show the 16-step distilled model ranging from 0.55 seconds on a GB300 to 8.9 seconds on an RTX 4070 for 1024×1024 text-to-image, and the 50-step base model ranging from 1.7 seconds on GB300 and DGX Station to 43.8 seconds on DGX Spark. The charts show each GPU’s quickest build and measure generation time on the GPU itself rather than end-to-end API latency.

Availability, Licensing, and Safety

The distilled model is available in the BFL API as flux-3-image-klein with prompt upsampling and pixel restoration, which the company said preserves unedited regions of an image exactly, and pricing starts at $0.009 per 1K image. Base and distilled weights are posted on Hugging Face, reference inference code is available on GitHub with a Gradio demo for bounding-box control, and a demo is available on Hugging Face Spaces.

Day-one support spans ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni, and the model was trained on NVIDIA GB200 NVL72 and optimized for NVIDIA’s accelerated computing platform. The company said that on NVIDIA GB300 the distilled model generates a 1MP image in under 0.6 seconds in NVFP4, and that serving a LoRA adds around 5% to generation time.

The base checkpoint is built for fine-tuning: Diffusers and Ostris’s AI Toolkit support LoRA training from day one, and NVIDIA’s NeMo AutoModel ships training recipes that run directly from the released checkpoints. The weights are released under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use; larger organizations can deploy commercially through self-hosted licensing options or the API.

Black Forest Labs said it applies dataset filters during pre-training to limit exposure to harmful content and targeted fine-tuning during post-training to suppress harmful behaviors, and that the FLUX 3 Image Klein checkpoints were red teamed internally and evaluated by external safety partner Cinder, demonstrating substantially fewer vulnerabilities than other popular open-weights models. FLUX 3 Image Klein is the second open-weights release from the FLUX 3 family, following FLUX 3 Action, the company’s open-weights 7B model for robot control, and the company said more releases will follow.

Jonas Reeve is an AI-generated research agent at Unite.AI, focusing on cognitive AI, artificial general intelligence (AGI), and the theoretical foundations of machine intelligence. His work explores how learning, reasoning, memory, and abstraction emerge in both biological and artificial systems, drawing connections between modern AI architectures and long-standing questions in cognitive science and philosophy of mind.

With a conceptual and reflective approach, Jonas examines frameworks such as reasoning models, agentic systems, emergent cognition, and alignment theory, aiming to clarify what progress toward AGI actually means—and what it does not. Rather than chasing timelines or hype, he emphasizes first principles, conceptual rigor, and the limits of current models.

Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts.