FLUX 3 Image Klein: 8 GB memory requirement and a free license under $5 million
Black Forest Labs released the open-weights model FLUX 3 Image Klein, at 5.6 billion parameters, on October 15, 2026, with day-one support in ComfyUI and Diffusers.
Black Forest Labs released the open-weights model FLUX 3 Image Klein, at 5.6 billion parameters, on October 15, 2026, with day-one support in ComfyUI and Diffusers. The INT8 version is said to run on GeForce RTX cards with 8 GB of memory, and the license allows free commercial use for businesses with under $5 million in annual revenue — but several of the numbers are the company's own, and the fine print is worth reading.
What was launched
The German AI research company Black Forest Labs launched FLUX 3 Image Klein on October 15, 2026 and released the model weights at the same time, according to Unite.AI and Crypto Briefing. The model has 5.6 billion parameters and covers three tasks: text-to-image generation, image editing, and generation with multiple references, supporting up to 10 reference images. It also supports bounding-box control and output up to 4K.
Base and distilled weights are on Hugging Face, reference code is on GitHub with a Gradio demo for bounding-box control, and there is a demo on Hugging Face Spaces. The day-one support list covers ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni. The model was trained on NVIDIA's GB200 NVL72 and optimized for NVIDIA's accelerated platform.
Hardware requirements and local performance
The weights require 11.2 GB of memory in BF16 and 5.6 GB in FP8. With the INT8 build, the model is said by the company, as summarized by Unite.AI, to run on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM via ComfyUI or Diffusers.
The company's own performance figures: the distilled model with INT8 is reported to generate a 1-megapixel image in around 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 via ComfyUI.
VentureBeat adds important caveats: Black Forest Labs has not tied the 8 GB limit to a specific model variant, output size, or generation time, and the speed charts accompanying the launch test neither an 8 GB configuration nor 4K output. In other words: the 8 GB requirement is real, but what actually works on such a card — and how fast — is unclear in the company's own published materials.
Licensing: free under $5 million
The weights are released under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5 million, in addition to research and non-commercial use. Larger organizations must use self-hosted licensing or the API.
There is, however, an open detail. VentureBeat points out that the company's published Kommunity license for FLUX 3 Action defines eligibility based on gross annualized revenue combined with affiliated companies — which would exclude a small subsidiary of a large corporation. The same license text also sets conditions around reviewing or filtering outputs and around redistribution. That text provides context for the license family, but the terms accompanying Klein's download are the controlling document — and they have not been confirmed beyond the $5 million threshold.
Benchmarks and API — with methodological caveats
Black Forest Labs claims, based on internal evaluations, that Klein ranks first in text-to-image, multi-reference generation, and image editing, ahead of larger open-weights models and above the closed models in the comparison, Ideogram V4.5 and FLUX.2 pro. Compared with the company's own FLUX.2 klein 9B, it is credited with a gain of 91 Elo in text-to-image and 70 in image editing, with 38 percent fewer parameters. These figures are the company's own internal assessments, not independently verified.
VentureBeat also flags a methodological skew in the company's API latency comparison: the reported median latency is 4.54 seconds for text-to-image and 7.18 seconds for editing at 1MP, measured from request to finished image including queue time — but Klein was measured on the company's own servers, while competitors were measured from the client, including polling and network time. That favors Klein.
The distilled model is available in BFL's API as flux-3-image-klein, with prompt upsampling and "pixel restoration," which the company says preserves unedited areas of an image exactly. Pricing per image, according to the API documentation as summarized by VentureBeat: $0.008 at 768×768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K, and $0.036 at 4K. Generation and editing cost the same at each resolution. No standard self-hosting fee for commercial use above the threshold has been published.
Safety claims
Black Forest Labs states that dataset filters are used during pretraining to limit exposure to harmful content, with targeted fine-tuning afterward. The company says Klein's checkpoints were "red team" tested internally and evaluated by the external security partner Cinder, with substantially fewer vulnerabilities than other popular open-weights models. These are the company's own claims, not independently verified.
What to check before deploying
For developers and small businesses weighing local deployment against the API, three things remain unclear in the launch materials: which variant, resolution, and speed the 8 GB limit applies to; whether the speed charts' omission of 8 GB and 4K testing reflects real performance on cheaper cards; and the exact license terms — particularly whether affiliated companies count toward the revenue threshold, and which conditions apply to filtering and redistribution. The model card and the license document accompanying the download are the place to verify this.