FLUX.3 Image Klein in ComfyUI: Files, VRAM and the $5M Licence Rule

Black Forest Labs released FLUX.3 Image Klein on 13 October 2026, the first open-weights image model from its FLUX.3 family (BFL’s earlier open FLUX.3 release, in September, was a set of robotics models). A day later its ComfyUI files went up, and they passed 338,000 downloads within four days. It is a 5.6B-parameter model that does text-to-image, editing with up to 10 reference images and layout with bounding boxes in one checkpoint, at resolutions up to 4K.

Two things matter before you download it. It is designed to be used with a separate 9B prompt-rewriting model, and its new licence, the FLUX Kommunity License, lets you use the images commercially only if your organisation earns under US$5 million a year. This guide covers both, along with the files, memory and ComfyUI settings. BFWR has not run the model; memory figures are BFL’s own measurements or our calculated estimates, and are labelled as such.

Quick answer

• ComfyUI support: native, in the stable release v0.41.0 (15 October 2026), with four official local templates.
• Download: about 17.7GB for the files the official template uses.
• Hardware: BFL says its INT8 files run on 8GB cards such as the RTX 3070. Comfy’s template recommends the BF16 weights only above 24GB.
• Prompt upsampler: BFL says not to use the model without it. ComfyUI’s templates have it switched on.
• Licence: free for non-commercial use. Commercial use of the images is allowed only below US$5M annual revenue, with conditions. Commercial fine-tuning or serving the model needs a paid licence.

What FLUX.3 Image Klein is

BFL’s reference repository describes it as “the 5.6B open weights image model” from Black Forest Labs, covering “text-to-image, editing with up to 10 reference images, and layout control with bounding boxes, through one model”. The weights confirm the size: the distilled transformer is 11.10GB at BF16, or 5.55 billion parameters. The text encoder is Qwen3-VL 4B (the same encoder Krea 2 uses), and the autoencoder is FLUX.2’s.

There are two checkpoints:

BFL’s README illustrates the speed gap on one H200 at 1 megapixel: “about 3 s against 12 s per image”. Comfy-Org’s own model page adds that, in BFL’s internal evaluations, it “improves on FLUX.2 [klein] 9B” while using fewer parameters. Those are vendor evaluations, not independent tests.

The prompt upsampler is not optional

The model was trained on long, detailed captions, and BFL is unusually direct about it: “Do not use either model without the prompt upsampler.” The README adds that “a prompt sent as typed gives noticeably weaker results.”

The upsampler is a separate 9B vision-language model fine-tuned from Qwen3.5-9B, released under Apache 2.0. Its model card explains why: “FLUX 3 Image Klein is not meant to run without a prompt upsampler in front of it”. It turns a short request such as “A man walking with his dog, afternoon sun” into a paragraph describing lighting, framing, clothing and position. In ComfyUI it loads through Load CLIP with the type set to flux3, and the templates expose an enable_prompt_upsampler switch.

That has a memory cost, because the upsampler is bigger than the image model. If they do not fit on your card together, BFL’s own advice for its reference code is to run the upsampler first, save the caption, and then generate. ComfyUI loads and unloads models as it goes, which helps on smaller cards, at the price of time.

ComfyUI support and templates

Native support arrived in ComfyUI pull request #17064, “Add Flux3-image model support”, and is included in the tagged v0.41.0 release from 15 October 2026. Update ComfyUI to 0.41.0 or later and you have it. That differs from MOSS-Image, another new model released the same week, whose native support is still only in the nightly build (see our MOSS-Image guide).

Comfy-Org added four local templates on 15 October: text to image and image edit, each for the distilled and Base checkpoints. All four list ComfyUI 0.41.0 as the minimum. Do not confuse them with the FLUX 3 Image partner node templates. Comfy-Org’s own page is clear that the partner node “calls Black Forest Labs’ hosted API and runs the larger FLUX 3 Image model”, while Klein is the open model you run yourself. Our guide to ComfyUI templates explains how to tell local templates from API ones.

Files and download size

BFL publishes ComfyUI-ready files in black-forest-labs/flux-3-image-klein-comfy. Like the main repository, it is gated: you accept the licence on Hugging Face before downloading. Sizes from the file listing:

The distilled template’s four files come to about 17.7GB. The upsampler also comes as INT8 (9.75GB) and BF16 (18.82GB) files.

NVFP4 is NVIDIA’s 4-bit format for its Blackwell generation. BFL’s reference code notes that --nvfp4 “needs Blackwell”, so on older cards use the INT8 files.

How much memory it needs

BFL’s own measurement is the firmest number available: peak GPU memory reported by nvidia-smi on a B200 at 1024×1024, with the upsampler not running.

Those figures come from BFL’s reference code on a data-centre GPU, which keeps everything resident. ComfyUI is more frugal: BFL itself recommends ComfyUI’s INT8 files for “An 8 GB GPU (RTX 3070 class)”, and Comfy-Org’s template switches to BF16 only “With more than 24 GB of VRAM”.

For planning, here is what our image generation VRAM calculator estimates for the image model alone, from the file sizes plus an activation allowance per megapixel borrowed from FLUX.2. These are estimates, and they come out lower than BFL’s measured BF16 peak, so treat them as a floor and leave headroom:

The upsampler is not included in those numbers, and it is the largest model in the chain. On a 12GB card, expect ComfyUI to swap models in and out between the upsampler and generation steps. Our model page, FLUX.3 Image Klein VRAM requirements, shows the same estimates across every precision and offload mode.

Settings

For the Base checkpoint, BFL’s defaults are 50 steps with CFG 4. For editing, the first reference image sets the output size, and up to 10 references are accepted.

The licence: what the FLUX Kommunity License allows

This is the part most write-ups skip. FLUX.3 Image Klein is not under FLUX.2 [klein] 4B’s Apache 2.0, nor FLUX.1 [dev]’s non-commercial licence. It uses a new document, the FLUX Kommunity License v1.0. BFL’s copy is behind the Hugging Face gate. We read the text from a publicly mirrored copy, and its key threshold is confirmed by SGLang’s FLUX.3 documentation and by Comfy-Org’s own FLUX 3 Image Klein page. This is a reading of the text, not legal advice.

The opening sentence sets the scope: the weights are “freely available for your non-commercial and non-production use or commercial use as a Qualifying User”. A Qualifying User is defined as you, “if at any time you together with your affiliates’ gross annualized revenue is less than US$5,000,000.”

What a Qualifying User gets is narrow. The grant allows you to “use Outputs, including Outputs from Derivatives, for commercial purposes”, and to run the model in production “solely for the purpose of generating such Outputs”. In plain terms:

• Under US$5M revenue: you may sell or use the images commercially. You must either run content filters or “ensure Output undergoes review for unlawful or infringing content” before it is distributed, and label it as AI-generated “to the extent required under applicable law”.
• Not covered, at any size: distributing the model or a fine-tune commercially, or building it into a product for your own users, which is what BFL’s paid Platform tier is sold for. The licence’s own definition of non-commercial excludes use “to train, fine tune, or distill other models for commercial use”.
• Over US$5M revenue: non-commercial and non-production use only, unless you buy a licence. BFL sells tiers on its licensing page, from Builder upwards.
• For everyone: you may not use the model or its outputs “to improve any other artificial intelligence model that performs any function similar to a FLUX Model”, and the licence is described as “revocable”.

Comfy-Org summarises it the same way: “It depends on your revenue.” If you need an image model you can build a product on without these conditions, FLUX.2 [klein] 4B and MOSS-Image are both Apache 2.0. Our Flux LoRA licence guide covers how the earlier FLUX licences treat fine-tunes.

Fine-tuning and LoRAs

The Base checkpoint is the one to train on. Hugging Face’s FLUX.3 post includes a LoRA training example with diffusers, and Comfy-Org notes that LoRAs trained on Base load through ComfyUI’s standard LoRA loader. Under the Kommunity License, a fine-tune for your own non-commercial use is fine. A fine-tune for commercial use is not covered and needs a licence from BFL.

FAQ

Is FLUX.3 Image Klein free?

Free to download and use for non-commercial purposes. Commercial use of the images is free only for organisations under US$5M annual revenue, with the filtering and disclosure conditions above.

Will it run on an 8GB graphics card?

BFL says the INT8 files run on 8GB cards such as the RTX 3070 in ComfyUI. It will be slower there than on a larger card, because the 9B prompt upsampler and the image model have to take turns on the GPU.

Which ComfyUI version do I need?

0.41.0 or later, released 15 October 2026. The official FLUX.3 Klein templates list it as the minimum.

Distilled or Base?

Distilled for making images: 16 steps and no guidance. Base for training LoRAs or fine-tunes.

Can I skip the prompt upsampler?

You can, but BFL says results are “noticeably weaker”. If memory is the problem, run the upsampler on its own first and paste its caption in as the prompt.

Related guides and tools

• Qwen-Image 2.1 in ComfyUI, another new model with a restrictive licence
• How to run Stable Diffusion locally, including installing ComfyUI
• Best local image generation models in 2026, every model compared on memory and licence
• ComfyUI hub and image model directory
• Stable Diffusion Prompts
• AI Upscale Models in 2026

Sources

• black-forest-labs/flux-3-image-klein on GitHub: README, model overview, memory table and settings
• FLUX 3 Image Klein ComfyUI files and the prompt upsampler model card
• Hugging Face: FLUX.3 Image Klein
• ComfyUI pull request #17064 and the workflow templates
• FLUX Kommunity License v1.0, read from a public mirror; threshold confirmed by the SGLang FLUX.3 cookbook
• Black Forest Labs: licensing

Updated 18 October 2026. File sizes, licence text and ComfyUI versions were read on that date. Memory figures are BFL’s measurements or our estimates, as labelled.