--- title: "How to Run FLUX 3 Klein Locally in ComfyUI: VRAM and License" url: "https://goremy.ai/blog/articles/flux-3-klein-local-install-comfyui-vram-license" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-26T19:00:09.054Z" updated: "2026-10-26T19:00:09.054Z" topic: Open Model Run Guides description: "Run FLUX 3 Klein locally in ComfyUI: 5.6B transformer, 16-step distilled model, about 26.4GB VRAM at full precision, and the under-$5M commercial license." --- # How to Run FLUX 3 Klein Locally in ComfyUI: VRAM and License > Run FLUX 3 Klein locally in ComfyUI: 5.6B transformer, 16-step distilled model, about 26.4GB VRAM at full precision, and the under-$5M commercial license. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-26 ## How do you run FLUX 3 Klein locally? You run FLUX 3 Klein locally by loading its open weights into a ComfyUI workflow. The model has a 5.6 billion parameter image transformer and uses Qwen3 VL 4B as its text encoder. In one full-precision test, it used 26,418 MB of VRAM, just under 27 GB. The license allows non-commercial use, plus commercial output for qualifying users under $5 million in annual revenue. ## TL;DR - FLUX 3 Klein is an open-weights model from Black Forest Labs with a **5.6B parameter image transformer** that handles text-to-image, editing, and multi-reference generation in one model. - It accepts **up to 10 reference images**, and the prompt can place elements using bounding boxes. - The distilled version runs in **16 steps without guidance**, though one tester found it slower than the previous Klein release. - A full-precision run in ComfyUI used about **26.4 GB of VRAM**, so a 24 GB consumer card is unlikely to fit it without a smaller format. - The text encoder is **Qwen3 VL 4B**, which is a separate model that ComfyUI has to load alongside the transformer. - Commercial use is limited to **qualifying users under $5M in annual revenue**, so read the license before shipping anything paid. ## What is FLUX 3 Klein? FLUX 3 Klein is Black Forest Labs' newest open-weights image model in the Klein line. It is a single model that covers three jobs: text-to-image generation, image editing, and multi-reference generation. Earlier workflows often split these across separate models or pipelines. Three details define its design: - A 5.6 billion parameter image transformer. - Qwen3 VL 4B as the text encoder. - A distilled variant that needs only 16 steps and no guidance. The model can also render text inside images. Black Forest Labs itself has said text output might not be correct in this version, so treat that as a soft feature. ## What hardware does FLUX 3 Klein need? The only hard number available comes from a hands-on test in ComfyUI. With the model loaded at full precision, VRAM consumption was 26,418 MB. That is a bit less than the previous version used on the same setup, according to the tester. | Item | Value | |---|---| | Image transformer | 5.6B parameters | | Text encoder | Qwen3 VL 4B | | Distilled sampling | 16 steps, no guidance | | Reference images | Up to 10 | | VRAM (full precision, observed) | 26,418 MB (under 27 GB) | Some practical reading of that table: - The figure covers a full-precision load, so it is a ceiling for planning, not a minimum for every format. - A 24 GB card will likely need a reduced-precision or quantized version, if one is available. The sources do not confirm any such release, so check the official model page. - Rented GPUs are the usual fallback when local VRAM falls short. A 32 GB or larger card would cover the observed footprint with some headroom. ## How do you set it up in ComfyUI? The tester used a simple ComfyUI workflow, and the graph itself is short. The general process looks like this: 1. Install ComfyUI, or update an existing install so it supports the new model's nodes. 2. Download the FLUX 3 Klein weights from Black Forest Labs' official release and accept the license terms. 3. Download the Qwen3 VL 4B text encoder that the workflow expects. 4. Place each file in the folder ComfyUI expects for that model type. 5. Load a FLUX 3 Klein workflow (the tester published theirs in a GitHub repo) and point each loader node at the right file. 6. Set the distilled variant to 16 steps with no guidance, then run a prompt. Exact file names, folders, and node names depend on the release and your ComfyUI version, so follow the official instructions rather than guessing. If a node shows up red, the usual cause is an outdated ComfyUI install. ## What can the model do that earlier versions could not? The headline feature is multi-reference generation with up to 10 reference images. In one test, a few reference photos plus a short prompt (which was upsampled) produced a scene with a wooden lookout in foggy orange hills. The subject kept her hat, yellow dress, and long hair, and the lighting matched the background. A second test swapped the subject for the tester's own photo. The face stayed intact, the hat was removed, and the model added blue jeans that were not in the original. That is a clean example of reference-driven editing. Other features worth knowing: - **Bounding boxes in the prompt.** You can place elements in specific regions of the frame. - **In-image text.** A honey shop sign in English and Arabic came out crisp in English and acceptable in Arabic. Small jar labels were unreadable. - **Single-instruction edits.** Spreading jam on toast and adding blueberries worked. A hair color change kept the rest of the image intact. ## How good is the output? The tester's verdict was mixed. Prompt adherence looked improved. A flamenco dancer prompt landed the courtyard, bougainvillea, blue tiles, crimson ruffles, mantilla, and carnation. Lighting and reflections were consistently strong, including a Golden Temple water reflection and a hazy river market scene. The weak points were familiar: - Hands and arms still looked wrong in some shots, with an odd thumb bend and a half-hidden second hand. - A detailed instruction to fix a seatbelt on an AI-generated image failed. - Some scenes, like a Bavarian cattle drive, looked busy and plasticky. - Realism in edits like the toast image trailed the previous model, in the tester's view. - Generation was slower than the earlier Klein release. The tester concluded that text-to-image and image-to-image quality may be plateauing. That is one person's opinion from a small prompt set, not a benchmark. Run your own prompts before committing a pipeline to it. ## What does the FLUX 3 Klein license allow? The license permits non-commercial use. It also permits commercial use of generated output for qualifying users with under $5 million in annual revenue. Above that threshold, the terms as described do not cover you, and you would need to contact Black Forest Labs about a separate arrangement. Two cautions. First, the summary here comes from a secondhand description, so read the actual license text on the official model page. Second, "commercial output" and "commercial use of the model" can be different things in a license. If you plan to host the model as a paid service or fine-tune and redistribute it, confirm that those uses are allowed. ## Is FLUX 3 Klein worth running locally? It depends on your hardware and your goal. The all-in-one design is convenient if you want generation, editing, and multi-reference work from one model in one workflow. The 16-step distilled variant keeps iteration reasonable. If you already have a 32 GB class GPU or rent one, the setup cost is low. If you have a 24 GB card, the full-precision footprint is a problem. If photorealism is your top priority, early impressions suggest the previous Klein release may still look more natural in some cases. If you build commercial products, the revenue cap matters more than any speed or quality difference. ## Frequently Asked Questions ### How much VRAM does FLUX 3 Klein use? In a ComfyUI test at full precision, it used 26,418 MB, just under 27 GB. That was slightly less than the previous version. Smaller formats may reduce this, but the sources do not confirm any. ### Can I use FLUX 3 Klein commercially? Partly. The license allows non-commercial use and commercial output for qualifying users under $5 million in annual revenue. Larger companies should read the full license and likely seek separate terms. ### How many reference images does it support? Up to 10. You can feed several photos into one generation, and the same model also handles plain text-to-image and single-image edits. ### Does FLUX 3 Klein render text accurately? Sometimes. Short signs in English rendered cleanly in testing, and Arabic looked acceptable, but small text was unreadable. Black Forest Labs has said text might not be correct in this version. ### Which text encoder does it use? It uses Qwen3 VL 4B. This is a separate model from the 5.6B image transformer, so your workflow must load both.