The weights are out. 8GB is the INT8 path
On October 15, Black Forest Labs released FLUX 3 Image Klein. The machine time is 15:00 UTC, 11:00 p.m. in Beijing. It is a 5.6B-parameter open-weights model for text-to-image, image editing, and generation with up to 10 reference images. They say it shares capabilities with the full FLUX 3 Image model: native 4K, pixel-accurate editing, and bounding boxes that place elements or confine an edit to a region.
The release includes a base model for fine-tuning, a step-distilled model for low latency, and FP8, INT8, and NVFP4 builds made with NVIDIA support. The license is the FLUX Kommunity License. Generator weights are 11.2GB in BF16 and 5.6GB in FP8. The INT8 build runs on an NVIDIA GeForce RTX with as little as 8GB of VRAM, through ComfyUI or Diffusers. With the distilled model and INT8, a 1MP image takes about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070, through ComfyUI.
[1]
The lead is an internal test, and the clocks are not in the same place
They write that in internal evaluations Klein ranks first for text-to-image, multi-reference generation, and image editing in the comparison, and also rates above Ideogram V4.5 and their own FLUX.2 [pro]. At 5.6B it is the smallest open model in that comparison, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. Against FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in image editing, with 38 percent fewer parameters. Parameter counts are for the image generator, as each publisher reports them. Text encoders are not counted. Krea 2 does not do image editing, so it is not in that comparison.
API latency is end to end, from the request to a finished 1MP image, including queueing. Klein was timed on their servers. Other models were timed from the client, which includes polling and network time. Krea 2 was timed on its Medium tier. Its Large tier took 31.1 seconds for text-to-image. The latency claim is a result under that timing.
[1]$0.009 is the starting price per 1K image
The distilled model on the API is named flux-3-image-klein. It includes prompt upsampling and what they call pixel restoration, which they say keeps unedited regions exact. Pricing starts at $0.009 per 1K image. On an NVIDIA GB300, the distilled model generates a 1MP image in under 0.6 seconds in NVFP4. Serving a LoRA adds about 5 percent to generation time.
The license allows free commercial use for individuals and for companies with annual revenue under $5 million, plus research and non-commercial use. Larger organizations use their self-hosted license or the API. On safety, they write that the checkpoints were red-teamed internally and evaluated by their external partner Cinder, and that they showed substantially fewer vulnerabilities than other popular open-weights models. The post does not say how many fewer. This is the second open-weights release in the FLUX 3 family. The previous one is FLUX 3 Action, a 7B model for robot control. How that control works is not in this post.
[1]