Black Forest Labs · Released, not on our provider

FLUX 3 Image Klein Comfy (2026): what the ComfyUI repository holds, which file to pick, what GPU it needs and what the license allows

FLUX 3 Image Klein Comfy is Black Forest Labs’ gated Hugging Face repository that packages its 5.6B-parameter FLUX 3 Image Klein and Klein Base image models as single files in bf16, int8 and NVFP4 for ComfyUI, where the int8 builds run on GPUs with as little as 8 GB of video memory, under the FLUX Kommunity License.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not FLUX 3 Image Klein Comfy.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Black Forest Labs
Type
Image
Released
2026-10-15
OpenRouter id
None

Checked

FLUX 3 Image Klein Comfy: facts, with sources

  • Black Forest Labs’ blog post is dated October 15, 2026 and says it is releasing FLUX 3 Image Klein, a 5.6B parameter open weights model for text-to-image generation, image editing and multi-reference generation with up to 10 reference images. The release has a base model for fine-tuning and a step-distilled model for low latency, with FP8, INT8 and NVFP4 builds developed with NVIDIA support, under the FLUX Kommunity License. BFL calls it the second open weights release of the FLUX 3 family, after FLUX 3 Action.Source: bfl.ai · checked 2026-11-19
  • Hugging Face’s API shows the black-forest-labs/flux-3-image-klein-comfy repository was created on October 14, 2026, one day before the blog post, and last modified on November 2, 2026. It is gated with automatic approval, tagged for ComfyUI, with the license shown as “other” under the name flux-kommunity-license and three base models listed: flux-3-image-klein, flux-3-image-klein-base and flux-3-image-klein-pup. Without logging in and accepting the terms, its README and files cannot be read, so we could not read the repository’s own card.Source: huggingface.co · checked 2026-11-19
  • BFL’s GitHub README describes this repository in its “Run it elsewhere” table: the ComfyUI row points to flux-3-image-klein-comfy and says it holds bf16, int8 and NVFP4 files for both models, with the int8 files running on 8 GB GPUs such as the RTX 3070. The README adds that BFL’s own code has no int8 path, so on GPUs without FP8 (RTX 30 series and older) you use these int8 files in ComfyUI. That is what the “comfy” in the name means: ready-made single-file builds for ComfyUI rather than the Diffusers folders in the main repositories.Source: github.com · checked 2026-11-19
  • BFL’s docs list two variants, both 5.6B parameters with bounding-box support and output up to 4K. Klein (distilled) runs 16 inference steps with guidance (CFG) 1 and is meant for fast editing and generation and local deployment. Klein Base runs 50 steps with CFG 3.5 and is for fine-tuning, LoRA training and research. BFL’s GitHub README table instead gives Base “CFG 4”, so the two BFL pages differ on that one number. BFL says the model fits on consumer GPUs down to an RTX 3090.Source: docs.bfl.ai · checked 2026-11-19
  • On memory and speed BFL’s blog says the generator weights take 11.2 GB in BF16 and 5.6 GB in FP8, and that with the INT8 build the model runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. With the distilled model and INT8 it generates a 1 MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI (timings measured by NVIDIA), and in under 0.6 seconds in NVFP4 on an NVIDIA GB300. The prompt upsampler is a separate 9B model on top of these figures.Source: bfl.ai · checked 2026-11-19
  • BFL’s GitHub README gives peak GPU memory on a B200 at 1024x1024 without the upsampler: 31 GB in bf16, 26 GB with --fp8 and 25 GB with --nvfp4 for the distilled model, and 31, 29 and 26 GB for Base. It says --fp8 needs an Ada, Hopper or Blackwell GPU and NVFP4 needs a Blackwell GPU. On one H200 at 1 MP, distilled takes about 3 s per image and Base about 12 s. The README’s install was tested on a B200.Source: github.com · checked 2026-11-19
  • The prompt upsampler is a separate repository, black-forest-labs/flux-3-image-klein-pup: a 9B vision-language model fine-tuned from Qwen3.5-9B that turns a short request, up to 10 reference images and optional bounding boxes into one detailed prompt. BFL says to always run it with Klein, because the model is trained on the captions it writes; its output is in English, and it is released under Apache 2.0. The ComfyUI workflows have an enable_prompt_upsampler switch, and the upsampler file is 18.8 GB in bfloat16 on its own repository.Source: huggingface.co · checked 2026-11-19
  • BFL’s blog says the weights are under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use, and that larger organisations can deploy commercially through self-hosted licensing or the API. BFL’s model page is stricter in tone: running Klein in a product needs a commercial license that covers self-hosted deployment at scale, and the weights come with the license on their Hugging Face card, which you should read before use. The license text sits behind the gate and we did not read it.Source: bfl.ai · checked 2026-11-19
  • BFL’s licensing page sells paid tiers that include FLUX 3 Image Klein: Builder (10K images a month, 1 domain, 10 licensed users, fine-tuning and LoRA rights, not for client use or downstream applications), Platform (100K images a month for SaaS and product teams), Professional (agencies producing images for named clients, first 3 clients included) and Enterprise (custom terms). The page we read gives no prices, only Contact Sales and Purchase buttons.Source: bfl.ai · checked 2026-11-19
  • BFL’s GitHub README says safety checks are off by default in the reference code: with --safety, prompts, reference images and outputs are checked for copyrighted characters, brands and public figures (by a vision-language model) and for NSFW content (by a classifier), and anything flagged is refused. Saved images carry EXIF tags “Software: AI generated” and “Make: Black Forest Labs”.Source: github.com · checked 2026-11-19
  • BFL’s blog describes its safeguards as dataset filters during pre-training, targeted fine-tuning in post-training, internal red teaming, and an evaluation by its external safety partner Cinder, which BFL says found substantially fewer vulnerabilities than other popular open weights models. That is BFL’s statement, not our test.Source: bfl.ai · checked 2026-11-19
  • The weights are free to download once you accept the license. BFL’s hosted API serves the distilled model as flux-3-image-klein, with prompt upsampling and pixel restoration included, and the blog says pricing starts at $0.009 per 1K image; the docs price table did not load as text in our read, so we do not quote other sizes.Source: bfl.ai · checked 2026-11-19
  • ComfyUI’s docs say the FLUX 3 Klein templates need ComfyUI 0.41.0 or later (shipped in the 0.41.1 backport), and that you accept the license on the gated repository before downloading. The templates load the int8 (convrot) files to keep memory low; with more than 24 GB of VRAM ComfyUI suggests the bf16 build and the bf16 upsampler, which follow long prompts more closely. The loaders are Load Diffusion Model for the checkpoint, Load CLIP with type flux3 for the Qwen3-VL 4B text encoder, and Load VAE; the full template defaults to 50 steps, the distilled one to 16 steps with cfg 1.Source: docs.comfy.org · checked 2026-11-19 · second-hand
  • ComfyUI’s own FLUX 3 Image Klein page summarises the license for its readers: free non-commercial use; commercial use of generated images if you and your affiliates have under US$5M in annual revenue, provided you filter or review outputs and disclose AI generation where the law requires; and any other commercial use, including commercial fine-tunes, needs a license from Black Forest Labs. This is ComfyUI’s reading, not the license text.Source: comfy.org · checked 2026-11-19 · second-hand
  • On November 19, 2026 Hugging Face’s API lists the ComfyUI repository with 1,037,285 downloads and 3,485 likes, against 151,721 downloads and 1,052 likes for the Diffusers repository flux-3-image-klein. These are Hugging Face counters, not BFL statements.Source: huggingface.co · checked 2026-11-19 · second-hand
  • FLUX 3 Image Klein is not on OpenRouter’s model list (checked November 19, 2026; the list has no Black Forest Labs model), and this site only chats with text models, so it cannot run it. A text model such as Qwen3.8 Flash or Claude Haiku 5.5 can be chatted with here.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

FLUX 3 Image Klein Comfy: FLUX 3 Image Klein, Klein Base and the prompt upsampler side by side

  • BFL says that in its internal evaluations FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation and image editing among the open weights models compared, and above two closed models, Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B it gains 91 Elo in text-to-image and 70 in image editing with 38% fewer parameters; the others compared are Qwen-Image 2.1 (7B) and Krea 2 (12.9B). These are BFL’s own ratings, not tested by us.Source: bfl.ai · checked 2026-11-19

Maker’s numbers

FLUX 3 Image Klein Comfy: FLUX 3 Image Klein variants: steps, guidance, memory, speed and license

People who search for FLUX 3 Image Klein Comfy usually ask which file to download. The short version from BFL’s own pages: take the distilled Klein for fast editing and generation, take Klein Base only if you will fine-tune or train a LoRA, and always run the prompt upsampler with either. The ComfyUI repository holds both models, so the choice is made by the file name, not by the repository. For a free-licensed community image model on a different base, see MossYume.

Our table: figures as printed on BFL’s docs page and GitHub README, read on November 19, 2026; BFL’s own numbers, not run by us · checked 2026-11-19
ItemKlein (distilled)Klein BasePrompt upsampler
Parameters5.6B5.6B9B (vision-language)
Inference steps1650Not applicable
Guidance (CFG)None (docs: 1)3.5 in the docs, 4 in the READMENot applicable
Maximum outputUp to 4KUp to 4KWrites a prompt
Seconds per 1 MP image on one H200About 3About 12Not stated
Peak memory on a B200, bf16, 1024x1024, no upsampler31 GB31 GBReserves 0.35 of the GPU by default
Best forFast editing and generation, local useFine-tuning, LoRA training, researchRewriting a short request into a detailed prompt
LicenseFLUX Kommunity LicenseFLUX Kommunity LicenseApache 2.0

Source: docs.bfl.aiSource: huggingface.coSource: github.com

Our table

Which FLUX 3 Image Klein Comfy file for which GPU

  • The file list Hugging Face’s API gives for the repository: six diffusion_models files (flux3-image-klein and flux3-image-klein-base, each in bf16, int8-convrot and nvfp4), five text_encoders files (qwen3vl_4b in bf16 and int8_convrot, and a qwen3.5_9b_flux3_prompt_upsampler in bf16, int8_convrot and w4a8), the FLUX.2 VAE (vae/flux2-vae.safetensors), twelve ready-made workflow JSON files (base and distilled, bf16 and int8, each for text-to-image, image-to-image and multi-image-to-image), LICENSE.md and NOTICE.md. Hugging Face counts about 93.7 GB (93,726,613,295 bytes) stored for the repository.Source: huggingface.co · checked 2026-11-19

Maker’s numbers

FLUX 3 Image Klein Comfy: The three builds in the ComfyUI repository and what BFL and ComfyUI say about them

The three builds trade file size and memory for the hardware they need. The table keeps to what the pages say and leaves a cell as “not stated” where they do not.

Our table: file names from the Hugging Face API; hardware and timing notes from BFL’s blog and README and ComfyUI’s docs, read on November 19, 2026; timings were measured by NVIDIA and quoted by BFL · checked 2026-11-19
Itembf16int8 (convrot)nvfp4
Distilled file name in the repositoryflux3-image-klein-bf16.safetensorsflux3-image-klein-int8-convrot.safetensorsflux3-image-klein-nvfp4.safetensors
Generator weights size (BFL)11.2 GBNot statedNot stated
Hardware noteComfyUI suggests it above 24 GB of VRAMAs little as 8 GB of VRAM (RTX 3070 class); the ComfyUI templates default to itNeeds a Blackwell GPU (BFL README)
Speed example, distilled, 1 MP imageNot statedAbout 4 s on an RTX 5090, under 9 s on an RTX 4070 (ComfyUI)Under 0.6 s on an NVIDIA GB300

Source: huggingface.coSource: bfl.aiSource: docs.comfy.orgSource: github.com

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to FLUX 3 Image Klein Comfy you can use now

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

$0.10 / $0.50 (per million tokens)

  • Anthropic’s model page says Claude Haiku 5.5 was released on October 7, 2026, is the latest Haiku, and will not be retired sooner than October 7, 2027. Anthropic’s launch post is dated the same day.Source: platform.claude.com · checked 2026-11-17
  • In its launch post Anthropic calls Haiku 5.5 "the cheapest, fastest, and most capable small model we’ve ever released", for high-volume, cost-sensitive tasks such as summaries, compactions, database queries and classification, and as a coding subagent alongside Opus 5.5 and Sonnet 5.5. It calls Haiku 5.5 its first Haiku-class model with an adjustable effort setting.Source: anthropic.com · checked 2026-11-17
  • Anthropic’s model page says it is built for high-volume, latency-sensitive work such as classification, routing, extraction and subagent tasks, and rates its latency "Fastest" in the current lineup (Fable 5.2 and Opus 5.5 are "Moderate", Sonnet 5.5 is "Fast").Source: platform.claude.com · checked 2026-11-17

FAQ

FLUX 3 Image Klein Comfy FAQ

What is FLUX 3 Image Klein Comfy?

It is the Hugging Face repository black-forest-labs/flux-3-image-klein-comfy, which Black Forest Labs points to for running FLUX 3 Image Klein in ComfyUI. It holds bf16, int8 and NVFP4 single-file builds of both the distilled Klein and Klein Base, the text encoder, the VAE, the prompt upsampler and ready-made workflows. The model itself is the 5.6B-parameter FLUX 3 Image Klein released on October 15, 2026.

Which file should I download from it?

For fast editing and generation, take the distilled Klein; take Klein Base only to fine-tune or train a LoRA. On an 8 GB card use the int8 files, which are also what the ComfyUI templates load by default; ComfyUI suggests bf16 above 24 GB of VRAM, and NVFP4 needs a Blackwell GPU. You also need the Qwen3-VL 4B text encoder and the FLUX.2 VAE from the same repository.

Why is the repository gated, and how do I get in?

Hugging Face marks it gated with automatic approval: you log in, open the repository and accept the license agreement, and access follows. We could not read the repository’s own card without doing that, so this page relies on BFL’s blog, docs and GitHub README and on ComfyUI’s docs. Do not download the files from a copy on another account without checking the license.

What GPU do I need to run it?

BFL’s blog says the INT8 build runs on NVIDIA GeForce RTX cards with as little as 8 GB of VRAM through ComfyUI, and its docs say the model fits on consumer GPUs down to an RTX 3090. BFL’s own code peaked at 31 GB on a B200 in bf16 without the upsampler. The 9B prompt upsampler adds to memory, so BFL suggests running it on its own first if both do not fit.

Is FLUX 3 Image Klein free for commercial use?

Only within the license. BFL’s blog says the FLUX Kommunity License allows free commercial use for individuals and companies under $5M in annual revenue, plus research and non-commercial use; larger organisations need self-hosted licensing or the API. BFL’s model page says running Klein in a product needs a commercial license. The license text is behind the gate and we did not read it, so check it yourself.

What is the prompt upsampler, and do I need it?

It is a 9B vision-language model, black-forest-labs/flux-3-image-klein-pup, fine-tuned from Qwen3.5-9B under Apache 2.0, that rewrites a short request into the detailed description Klein was trained on. BFL says to always run it, because a prompt sent as typed gives noticeably weaker results. In the ComfyUI templates it is on by default and has an enable_prompt_upsampler switch.

Can I run FLUX 3 Image Klein on this site?

No. It is not on OpenRouter’s model list, and this site only chats with text models, so it cannot generate images. A text model can still help you write the detailed prompt Klein likes: on this site a typical reply to Qwen3.8 Flash costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Hy4 Flash is another small chat model here.

Is this the official Black Forest Labs website?

No. This page is unofficial and not affiliated with or endorsed by Black Forest Labs. Facts link to the pages we read.