Is FLUX 3 Klein free?

The weights are published under the FLUX Kommunity License v.1.0. The verified table does not list a download price. Non-commercial use is allowed, and a Qualifying User may use Outputs commercially if you together with your affiliates have gross annualized revenue under US$5,000,000. The hosted API is separate. The blog says: "Pricing starts at $0.009 per 1K image."

The weights are available, and the inference code repository is Apache-2.0. The weights themselves are under the FLUX Kommunity License, which is not an OSI open-source license. The text encoder, Qwen3-VL-4B-Instruct, and the official prompt upsampler are Apache-2.0. "Open weights" here means you can download them after accepting the gate, not that every use is unrestricted.

Can I use it commercially?

A Qualifying User may use Outputs commercially, including Outputs from derivatives, and may run the model in production for the purpose of generating those Outputs. Section 4.b prohibits using the model, derivatives, or Outputs to improve another model that performs a similar function. The model page FAQ says: "Running Klein in a product needs a commercial license." The plain-English page is not legal advice.

On a B200, at 1024×1024, with the upsampler off, the peaks are 31 GB in BF16, 26 GB in FP8, and 25 GB in NVFP4. The BF16 generator weights are 11.2 GB in the blog and 11.1 GB as a ComfyUI file. The official BF16 upsampler file is 18.8 GB, and on an 80 GB card that BF16 upsampler adds about 28 GB. Our own test used the w4a8 upsampler (6.7 GB file) and added about 5.8 GB of peak VRAM, not 28 GB.

The blog says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. The ComfyUI repository says: "The int8 files run on 8 GB GPUs such as the RTX 3070." By file size from the Hugging Face API, int8 generator 5.83 GB + int8 text encoder 4.86 GB + w4a8 upsampler 6.72 GB + VAE 0.34 GB is about 17.8 GB, so that full stack cannot stay resident on an 8 GB card. The official 8 GB figure must refer to running without the upsampler resident, or with offloading or sequential loading. We have not tested an 8 GB card. The docs also say the model "fits on consumer GPUs, down to an RTX 3090," and that conflict with the 8 GB claim is unconfirmed.

How fast is FLUX 3 Klein?

On our RTX 6000D, in ComfyUI 0.41.3, distilled Klein at 16 steps, CFG 1, and 1024×1024, warm images took about 5.9 s in FP8 and 5.8 s in BF16 with the upsampler off, and about 9.5 s and 9.6 s with the upsampler on. FP8 here means the BF16 file loaded as fp8_e4m3fn, not an official FP8 checkpoint. Peak VRAM for our process, with models kept resident on a shared GPU, was about 12.0 GB to 22.2 GB depending on the config. That peak is not a minimum requirement, and we did not test an 8 GB card.

Does ComfyUI support it?

Yes, natively, from ComfyUI v0.41.0 on 2026-10-15, with official templates. The Klein template needs ComfyUI 0.41.0 or newer, and 0.41.1 backported it. The distilled template defaults to 16 steps. ComfyUI files are BF16 11.1 GB, int8 5.8 GB, and NVFP4 5.3 GB, and NVFP4 needs a Blackwell GPU.

What is the difference between Klein and Klein Base?

Klein is distilled, defaults to 16 steps, uses no guidance, and is the checkpoint behind the Black Forest Labs API. Klein Base is undistilled, 50 steps, default CFG 3.5, for fine-tuning and LoRA. The GitHub README lists CFG 4 for Base. The ComfyUI tutorial describes the distilled template as 16 steps at CFG 1, adjustable from 4 to 16.

Do I need the prompt upsampler?

Yes, unless you expand the prompt another way the card allows. The model card says: "Do not use the model without a prompt upsampler." The official one is flux-3-image-klein-pup, a 9B model fine-tuned from Qwen3.5-9B, BF16 file 18.8 GB. Any LLM with the official system prompts is the other documented option. Diffusers will not run the upsampler for you.

How many reference images?

Up to 10. The model card, the GitHub README, and the Black Forest Labs blog all state that limit. The same sources list single-image editing as a separate capability, so one input image is the edit case and the cap of 10 is the multi-reference case. This site has not published its own count.

What is the max resolution?

4K. That ceiling is in the model card, the GitHub README, and the Black Forest Labs blog. It is the maximum they name, not a claim that every GPU in the memory table was measured at 4K. The B200 peaks of 31, 26, and 25 GB were measured at 1024×1024 with the upsampler off.

Yes. The Black Forest Labs API model id is flux-3-image-klein, and the blog says: "Pricing starts at $0.009 per 1K image." That is the publisher’s API. The distilled checkpoint behind it is the 16-step Klein model. This page does not list third-party hosts.

No. pixglen is an independent, unofficial resource. It is not affiliated with or endorsed by Black Forest Labs. FLUX is a trademark of its owner. The pages quote the publisher’s numbers and the license. They are not a product page, and they are not legal advice.

FLUX 3 Klein FAQ

FLUX 3 Klein is FLUX 3 Image Klein, a 5.6B image transformer from Black Forest Labs, released publicly on 2026-10-15. The short answers below use only the verified figures. They cover the license, memory, ComfyUI, the two checkpoints, reference images, resolution, the API, and whether this site is official. Longer pages linked from the related list go into the same facts with the sources attached.

Is FLUX 3 Klein free?: The weights are published under the FLUX Kommunity License v.1.0. The verified table does not list a download price. Non-commercial use is allowed, and a Qualifying User may use Outputs commercially if you together with your affiliates have gross annualized revenue under US$5,000,000. The hosted API is separate. The blog says: "Pricing starts at $0.009 per 1K image."
Is it open source?: The weights are available, and the inference code repository is Apache-2.0. The weights themselves are under the FLUX Kommunity License, which is not an OSI open-source license. The text encoder, Qwen3-VL-4B-Instruct, and the official prompt upsampler are Apache-2.0. "Open weights" here means you can download them after accepting the gate, not that every use is unrestricted.
Can I use it commercially?: A Qualifying User may use Outputs commercially, including Outputs from derivatives, and may run the model in production for the purpose of generating those Outputs. Section 4.b prohibits using the model, derivatives, or Outputs to improve another model that performs a similar function. The model page FAQ says: "Running Klein in a product needs a commercial license." The plain-English page is not legal advice.
How much VRAM?: On a B200, at 1024×1024, with the upsampler off, the peaks are 31 GB in BF16, 26 GB in FP8, and 25 GB in NVFP4. The BF16 generator weights are 11.2 GB in the blog and 11.1 GB as a ComfyUI file. The official BF16 upsampler file is 18.8 GB, and on an 80 GB card that BF16 upsampler adds about 28 GB. Our own test used the w4a8 upsampler (6.7 GB file) and added about 5.8 GB of peak VRAM, not 28 GB.
Does it run on 8 GB?: The blog says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. The ComfyUI repository says: "The int8 files run on 8 GB GPUs such as the RTX 3070." By file size from the Hugging Face API, int8 generator 5.83 GB + int8 text encoder 4.86 GB + w4a8 upsampler 6.72 GB + VAE 0.34 GB is about 17.8 GB, so that full stack cannot stay resident on an 8 GB card. The official 8 GB figure must refer to running without the upsampler resident, or with offloading or sequential loading. We have not tested an 8 GB card. The docs also say the model "fits on consumer GPUs, down to an RTX 3090," and that conflict with the 8 GB claim is unconfirmed.
How fast is FLUX 3 Klein?: On our RTX 6000D, in ComfyUI 0.41.3, distilled Klein at 16 steps, CFG 1, and 1024×1024, warm images took about 5.9 s in FP8 and 5.8 s in BF16 with the upsampler off, and about 9.5 s and 9.6 s with the upsampler on. FP8 here means the BF16 file loaded as fp8_e4m3fn, not an official FP8 checkpoint. Peak VRAM for our process, with models kept resident on a shared GPU, was about 12.0 GB to 22.2 GB depending on the config. That peak is not a minimum requirement, and we did not test an 8 GB card.
Does ComfyUI support it?: Yes, natively, from ComfyUI v0.41.0 on 2026-10-15, with official templates. The Klein template needs ComfyUI 0.41.0 or newer, and 0.41.1 backported it. The distilled template defaults to 16 steps. ComfyUI files are BF16 11.1 GB, int8 5.8 GB, and NVFP4 5.3 GB, and NVFP4 needs a Blackwell GPU.
What is the difference between Klein and Klein Base?: Klein is distilled, defaults to 16 steps, uses no guidance, and is the checkpoint behind the Black Forest Labs API. Klein Base is undistilled, 50 steps, default CFG 3.5, for fine-tuning and LoRA. The GitHub README lists CFG 4 for Base. The ComfyUI tutorial describes the distilled template as 16 steps at CFG 1, adjustable from 4 to 16.
Do I need the prompt upsampler?: Yes, unless you expand the prompt another way the card allows. The model card says: "Do not use the model without a prompt upsampler." The official one is flux-3-image-klein-pup, a 9B model fine-tuned from Qwen3.5-9B, BF16 file 18.8 GB. Any LLM with the official system prompts is the other documented option. Diffusers will not run the upsampler for you.
How many reference images?: Up to 10. The model card, the GitHub README, and the Black Forest Labs blog all state that limit. The same sources list single-image editing as a separate capability, so one input image is the edit case and the cap of 10 is the multi-reference case. This site has not published its own count.
What is the max resolution?: 4K. That ceiling is in the model card, the GitHub README, and the Black Forest Labs blog. It is the maximum they name, not a claim that every GPU in the memory table was measured at 4K. The B200 peaks of 31, 26, and 25 GB were measured at 1024×1024 with the upsampler off.
Is there an API?: Yes. The Black Forest Labs API model id is flux-3-image-klein, and the blog says: "Pricing starts at $0.009 per 1K image." That is the publisher’s API. The distilled checkpoint behind it is the 16-step Klein model. This page does not list third-party hosts.
Is pixglen official?: No. pixglen is an independent, unofficial resource. It is not affiliated with or endorsed by Black Forest Labs. FLUX is a trademark of its owner. The pages quote the publisher’s numbers and the license. They are not a product page, and they are not legal advice.

Questions

Sources

Related guides