Black Forest Labs releases 5.6-billion-parameter image model with open weights on Hugging Face
The German company announced on October 15, 2026 that the weights are now openly available, with quantized variants that reportedly run on consumer GPUs with as little as 8 GB of VRAM. But all performance and safety figures behind the release come from the company's own evaluations.
Black Forest Labs (BFL), based in Freiburg, announced on October 15, 2026 the open-weights release of FLUX 3 Image Klein, which the company itself describes as its "most capable open-weights image model to date" (2e59a8ab). The release means anyone can download the model, run it locally and fine-tune it — but it also comes with a raft of company-generated claims about performance and safety that have not yet been independently confirmed.
What the model is
FLUX 3 Image Klein is a text-to-image model with 5.6 billion parameters that, according to BFL, supports three modes of use: text-to-image generation, image editing, and multi-reference generation with up to 10 reference images (09e1b445). The company claims the model retains features from the larger FLUX 3 Image model, such as native 4K generation, bounding-box control, and "pixel-perfect" editing (2e59a8ab).
The release includes a base model intended for fine-tuning, a distilled model for low-latency generation, and quantized versions in FP8, NVFP4 and INT8 formats (2e59a8ab).
Hardware requirements
According to Unite.AI (09e1b445), the generator requires 11.2 GB of memory in BF16 and 5.6 GB in FP8. With the INT8 build, the model is said to run on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM via ComfyUI or Diffusers — opening up local image generation on a range of consumer machines that previously lacked the capacity to run large image models locally.
Availability and license
The base and distilled weights are on Hugging Face, with reference inference code on GitHub that includes a Gradio demo for bounding-box control. A demo is also available on Hugging Face Spaces (09e1b445). The distilled model is available through BFL's API as flux-3-image-klein, with pricing from $0.009 per 1,000 images.
The weights are released under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue under $5 million. Larger organizations can use the model commercially through self-hosted licensing arrangements or via the API (09e1b445).
Performance and safety claims — from BFL only
All performance figures tied to the release are the company's own. BFL reports that FLUX 3 Image Klein ranks first in internal evaluations in text-to-image, multi-reference generation and image editing among open-weights models, and that it "ranks above" the closed comparison models Ideogram V4.5 and FLUX.2 pro (09e1b445). The company also states that the model achieves 91 Elo points higher than its predecessor FLUX.2 klein 9B in text-to-image, and 70 points higher in image editing, with 38% fewer parameters — figures also reproduced by Crypto Briefing (caf347cc).
BFL reports a median latency of 4.54 seconds for text-to-image and 7.18 seconds for editing at 1 megapixel, measured from request to finished image including queue time. But according to Unite.AI, the methodology is not symmetric: FLUX 3 Image Klein was measured on BFL's own servers, while the competitor models were measured from the client side, with polling and network time included (09e1b445). The comparison is therefore not directly equivalent, and latency differences between models should be interpreted with caution.
On safety, BFL says the datasets are filtered during pre-training to limit exposure to harmful content, and that post-training fine-tuning is intended to suppress harmful behavioral patterns. The company claims the weights were internally "red team" tested and assessed by external security partner Cinder, with "substantially fewer vulnerabilities" than other popular open-weights models (09e1b445). No independent audit of these safety claims is available.
Developer side and support
The press release states that the model is supported in Diffusers, ComfyUI, vLLM-Omni and SGLang Diffusion (2e59a8ab). Worth noting is that the details of the Diffusers integration — for example, which version is required, or which API changes accompany it — are not documented in the available sources, all of which stem from the release materials.
BFL founder and CEO Robin Rombach ties the release to the company's broader strategy: "By sharing our research openly, we created a global community of developers and creators working together to advance the field," Rombach says. "FLUX 3 lets this community adapt the same foundation to very different needs, from creative generation to robotics" (2e59a8ab).
Open questions
So far, there are no independent measurements of the model — of image quality, latency or safety. The company's claims about Elo rankings, comparisons with Ideogram V4.5 and FLUX.2 pro, and "substantially fewer vulnerabilities" are so far documented only through BFL's own evaluation methods and press materials. Both Unite.AI and Crypto Briefing base their articles on the same release materials, so their agreement does not constitute independent confirmation. It remains to be seen whether the model holds up in independent testing, and how it performs in practice on hardware with 8 GB of VRAM — the most accessible, but also the most demanding configuration.