FLUX 3 Image Klein: Black Forest Labs' 5.6B open-weight image model, and what 'free under $5M' really covers

On October 15, 2026, German image-model lab Black Forest Labs (BFL) released FLUX 3 Image Klein, a 5.6-billion-parameter open-weight model that generates and edits images, including at 4K resolution, on a single consumer GPU. It's a small sibling of FLUX 3 Image, the paid model BFL launched on October 1. The weights are free to download, and small businesses can use the model commercially. The license is narrower than "free for companies under $5 million" suggests, though: it covers generating images, not distributing or hosting the model.

This story is based on BFL's launch post, model page, API release notes, Hugging Face model card and license file, and VentureBeat's reporting, checked October 16, 2026.

The short version

• What: a 5.6B image model for text-to-image, instruction-based editing, bounding-box placement and up to 10 reference images. A separate Qwen3-VL-4B text encoder and an optional "Pup" prompt-rewriting model come alongside it.
• Versions: a base model for fine-tuning and a step-distilled model for speed, with FP8, INT8 and NVFP4 builds.
• Speed (vendor-reported): about 4 seconds per 1MP image on an RTX 5090, under 9 seconds on an RTX 4070, from 8 GB of VRAM.
• License: FLUX Kommunity License v1.0. Commercial output generation is free if your company and affiliates earn under $5 million a year.
• API: from $0.008 to $0.036 per image, depending on resolution.

What is FLUX 3 Image Klein?

BFL calls Klein "the most capable open weights image model," a claim based on its own evaluations. Its launch post says Klein "shares core capabilities with the full FLUX 3 Image model," including native 4K generation, pixel-precise editing and bounding boxes that place objects or confine edits to a region.

The 5.6B figure covers the image transformer only. The model card lists the text encoder, Qwen3-VL-4B-Instruct (Apache 2.0), separately, and BFL's command-line tool runs the Pup prompt upsampler by default. Pup is about 18.8 GB and adds about 28 GB of GPU memory, per the model card, so turn it off (--no-pup) on smaller cards. Our guide to choosing a model for your laptop explains why parameters and memory don't map one to one.

How fast is it, and what does it cost?

All speed figures here are vendor-reported. BFL says Nvidia measured them, with RTX results in ComfyUI using the distilled model and the INT8 build:

Weights are 11.2 GB in BF16 and 5.6 GB in FP8, per BFL. VentureBeat notes the 8 GB minimum isn't tied to a resolution or speed, and that time rises steeply with size: on an RTX 4070, 16-step generation went from 8.9 seconds at 1K to 50.4 seconds at 2K. No timing was given for 4K on consumer cards.

BFL's API prices Klein per image, the same for generation and editing:

Last verified: October 16, 2026 (BFL API release notes; Klein also has a 1.5K tier at $0.011)

What does the FLUX Kommunity License actually allow?

This is the part to read carefully. The license file on Hugging Face, the FLUX Kommunity License v1.0, works like this:

• Who qualifies. You're a "Qualifying User" if "your affiliates' gross annualized revenue is less than US$5,000,000." The threshold counts affiliates, so a small team inside a large company doesn't qualify. The model card states the same threshold.
• What qualifying users can do commercially. Use outputs "for commercial purposes," and run the model and derivatives "in a production environment solely for the purpose of generating such Outputs" (§2.a). That covers, say, a small studio making paid client artwork or an app generating images for its users.
• What they can't. Distribution of the model is licensed "solely" for non-commercial purposes (§2.a). The model card says the exception isn't blanket permission for commercial fine-tuning or distribution, and using the model to "train, fine tune, or distill other models for commercial use" falls outside the non-commercial definition.
• Obligations. Qualifying users must run content filtering or review outputs for unlawful content, and add AI disclosures where the law requires them (§2.e).
• No training rivals. You can't use the model or its outputs "to improve any other artificial intelligence model that performs any function similar to a FLUX Model" (§4.b).
• Termination. The license ends automatically on any breach (§8.a), and BFL says it "may terminate this License, in whole or in part, at any time upon notice" (§8.b). Delaware law governs.
• Above $5 million. You need a commercial license. BFL's licensing page lists Builder, Platform, Professional and Enterprise tiers, with prices through sales.

BFL's model page words the limit more broadly than the license does: "Running Klein in a product needs a commercial license, which covers self-hosted deployment at scale." The license text itself lets qualifying users run the model in production to generate outputs. BFL claims no ownership of outputs. This is a summary of the license text, not legal advice.

Compared with the licenses in our open-weight license comparison, this is more restrictive than Apache or MIT and stricter in kind than revenue gates such as Meta's new Muse Model License, which permits distribution and resale and gates only large model-hosting businesses. The termination-on-notice clause means the free grant isn't irrevocable.

How does it compare with other image models?

BFL's internal evaluations rank Klein first among open-weight models for text-to-image, multi-reference generation and editing, and above the closed Ideogram V4.5 and FLUX.2 [pro]. BFL says Klein gains 91 Elo points in text-to-image and 70 in editing over its own FLUX.2 [klein] 9B, with 38% fewer parameters.

These are vendor-reported. VentureBeat notes that BFL's charts give no prompts, judging criteria, sample sizes or uncertainty, and that the comparison set leaves out OpenAI's GPT Image 2.5 and Google's Nano Banana 2.1. BFL also says safety partner Cinder evaluated the checkpoints, without publishing numbers.

Where can you run it?

Weights are on Hugging Face (black-forest-labs/flux-3-image-klein and -base), with day-one support in ComfyUI, Hugging Face Diffusers, SGLang Diffusion and vLLM-Omni, plus Nvidia fine-tuning recipes. Klein isn't a language model, so tools like Ollama don't apply here; ComfyUI and Diffusers are the main local paths.

What's confirmed, reported and uncertain

• Confirmed (BFL, Hugging Face): the October 15 release; 5.6B parameters; base and distilled versions and quantized builds; API prices; license terms quoted above.
• Vendor-reported: "most capable open weights image model," Elo gains, comparisons with closed models, speed and VRAM figures, and the Cinder safety review.
• Reported: VentureBeat's caveats on BFL's benchmarks and its RTX 4070 timings by resolution.
• Uncertain: self-hosted license prices above $5 million; independent quality rankings; real 4K speed on consumer GPUs.