You've got an RTX card with 8GB and you read that Black Forest Labs' new open image model runs on it. It does, in a sense. FLUX 3 Image Klein, published on 15 October, is a 5.6B image generator under the FLUX Kommunity License, and BFL's launch post says its INT8 build "runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM". What that line leaves out is the 9B prompt upsampler that BFL's own model card says Klein isn't meant to run without.
The short answer
Klein is two gated Hugging Face repos (a step-distilled model and an undistilled base for fine-tuning), plus a gated ComfyUI pack and an ungated, Apache 2.0 prompt upsampler. The 8GB figure is for the INT8 generator in ComfyUI. The INT8 files for the generator and its Qwen3-VL-4B text encoder already add up to 10.69 GB, and the upsampler is another 9.75 GB in INT8 or 6.72 GB in W4A8. The model card's own peak with the reference code is 25 GB in NVFP4 with the upsampler switched off. Commercial use of outputs is free only below $5M in annual revenue, affiliates included.
Where the 8GB comes from
We read the file lists through the Hugging Face API, and the INT8 build isn't where most people will look. The main flux-3-image-klein repo, the one Diffusers loads, ships BF16 weights (11.10 GB for the transformer) plus two quantised variants, fp8r at 6.27 GB and nvfp4 at 5.28 GB. No INT8. The INT8 files live in a separate repo, flux-3-image-klein-comfy, as int8-convrot safetensors: 5.83 GB for the distilled generator and 4.86 GB for the Qwen3-VL-4B text encoder. That's the build BFL means when it quotes 4 seconds for a 1MP image on an RTX 5090 and under 9 seconds on an RTX 4070, both "through ComfyUI".
Add those two files. 10.69 GB. That's more than 8GB before a single activation, so on an 8GB card the encoder and the generator can't both sit in VRAM at once. Our guess is that ComfyUI loads one, runs it, and swaps in the other, which works but costs time on every prompt. BFL doesn't say how the 8GB run is done, and we haven't run it ourselves, so treat that as arithmetic, not a measurement.
The other published numbers don't line up with 8GB either. The model card gives peak GPU memory with BFL's reference code on a B200 at 1024x1024, upsampler off: 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4. Hugging Face's Diffusers launch post measures lower, 20.72 GB peak for BF16 on an H100 and 15.3 GB for NVFP4 on an RTX 6000. Different code, different GPUs, so we wouldn't average them. And the card says it plainly: a quantised file's size "is not a measure of total GPU memory use".
The upsampler nobody counts
Here's the part that changes the hardware plan. BFL also published flux-3-image-klein-pup, a 9B fine-tune of Qwen3.5-9B that rewrites a short request into the long, detailed prompt Klein expects. Its model card doesn't hedge: Klein "is not meant to run without a prompt upsampler in front of it", and a short prompt passed as typed "gives generic output". The reference inference code loads it by default. The Klein card puts its cost at about 28 GB on an 80 GB GPU.
The upsampler is 18.82 GB in BF16. The ComfyUI pack has smaller builds, 9.75 GB in INT8 and 6.72 GB in W4A8, behind an enable_prompt_upsampler switch in the text-to-image templates. So the honest budget for the setup BFL recommends is a generator, a 4B text encoder and a 9B language model. On an 8GB card that's three models taking turns.
There's a way out, and it's a decent one. The upsampler card says any other LLM can do the job with the system prompts shipped in system_prompts.json. If you already run a hosted or local model for something else, point it at those prompts and keep the GPU for Klein. We think that's the setup most small teams should start with, though we'd check output quality against the stock upsampler before committing; BFL says its own is tuned for the job, and we haven't compared them.
The licence and the gate
All three Klein repos with the image model in them are gated. It's an automatic click-through, not a manual review: you agree to share your contact information and the download starts. That still means a Hugging Face token in your CI or your container build, which is worth knowing before you script a pull. The upsampler repo isn't gated and carries Apache 2.0.
The weights themselves sit under the FLUX Kommunity License v1.0, the same one BFL used for FLUX 3 Action base in September. The card's summary: non-commercial use, plus commercial output generation by qualifying users "with less than US$5M in gross annualized revenue, including affiliates". Above that, BFL sends you to its commercial licensing page. Two clauses in the v1.0 text matter for ops teams. Commercial use as a qualifying user requires you to "implement and maintain content filtering measures", so a bare Klein endpoint isn't compliant even under $5M. And the restrictions bar using Klein's output "to improve any other artificial intelligence model" that does what FLUX does, which rules out a synthetic image dataset for training another image model, at any revenue.
One thing we can't tell you yet is whether Klein is the "FLUX 3 Dev" BFL promised for "later this year" back in July. The Klein post doesn't mention Dev at all. We've kept Dev as a separate, still-promised line in our open-weight models tracker, where Klein now has its own row.
FAQ
Does FLUX 3 Image Klein run on an 8GB GPU?
BFL says the INT8 build does, through ComfyUI, on GeForce RTX cards with as little as 8GB. The INT8 generator and text encoder files total 10.69 GB, so they can't both be resident on such a card, and the 9B prompt upsampler BFL recommends is extra. The model card's reference-code peak is 25 GB in NVFP4 with the upsampler off.
Do I need the prompt upsampler?
BFL's upsampler card says Klein isn't meant to run without one, and that short prompts give generic output. You can use BFL's 9B upsampler or any other LLM with the system prompts in system_prompts.json.
Can I use Klein's images commercially?
Only if you're a qualifying user with less than $5M in gross annualized revenue, affiliates included, and you keep content filtering in place. Non-commercial and research use is allowed at any size. Larger companies need a commercial licence from BFL.
Which Klein checkpoint should I download?
The distilled flux-3-image-klein runs in 16 steps without guidance and is meant for generation and editing. The undistilled flux-3-image-klein-base uses 50 steps and is meant for fine-tuning and LoRA training. Both are 5.6B parameters, excluding the text encoder.
Sources: BFL launch post, Klein model card, Klein base model card, prompt upsampler model card, ComfyUI pack, Hugging Face model API, Hugging Face Diffusers launch post.
