FLUX 3 Image Klein: BFL's 5.6B open-weight image model, the US$5M licence line and the hardware it really needs
BFL's new open-weight image model runs on consumer GPUs, but free commercial use stops at US$5M revenue and the model card's memory figures are far higher.
Black Forest Labs (BFL) released FLUX 3 Image Klein on 15 October: an open-weight image model for generation and editing that it says runs on a single consumer GPU. For Indian agencies, D2C brands and product teams paying per image to closed APIs, a self-hostable model is worth a look. Before you build on it, though, check three things: the licence, the real memory footprint, and how much of the quality claim is BFL’s own testing.
What was released
According to BFL’s announcement, Klein is a 5.6B-parameter model that does text-to-image, image editing and multi-reference generation from up to 10 reference images. It shares three features with the larger FLUX 3 Image: native 4K output, “pixel-perfect editing”, and bounding boxes that control where elements go or limit an edit to one area. It is the second open-weight model in the FLUX 3 family, after FLUX 3 Action.
There are two checkpoints:
- Klein (step-distilled) for low-latency generation. The model card recommends 16 steps and no guidance.
- Klein Base for fine-tuning and LoRA training, at 50 steps and guidance 3.5.
BFL also ships FP8, INT8 and NVFP4 builds, made with NVIDIA. The text encoder is Qwen3-VL-4B-Instruct, which is a separate component. The 5.6B figure is the image transformer alone; Hugging Face’s own metadata shows 6B.
The licence: free commercial use stops at US$5M revenue
This is not an Apache-style release. The weights come under the FLUX Kommunity License v1.0 (BFL’s spelling). Per the model card:
- Non-commercial use is allowed.
- “Qualifying Users” with less than US$5M in gross annualised revenue, including affiliates, may generate commercial outputs, subject to the licence terms.
- That exception is not blanket permission for commercial fine-tuning or for distributing the model. Other commercial use needs a licence from BFL or goes through its API.
- Section 4 lists uses that are out of scope, including military purposes, surveillance and biometric processing.
- Downloads are gated. You must log in and agree to share contact details.
The text encoder is Apache-2.0, but that doesn’t change the terms for the image model.
What it means in practice: a small studio or early-stage startup in India will usually sit under the US$5M line, so generating marketing or product images is likely fine. Two cases need more care. One is a company that is part of a larger group, because affiliates count. The other is anyone planning to fine-tune Klein and ship it inside a product. Read the licence text itself, and get advice if you are near the threshold.
Hardware: two very different numbers
The headline and the model card don’t describe the same setup, so look at both.
BFL’s blog says the generator weights are 11.2 GB in BF16 and 5.6 GB in FP8, and that the INT8 build runs on GeForce RTX GPUs with “as little as 8GB of VRAM” through ComfyUI or Diffusers.
The model card measures 1024×1024 generation on an NVIDIA B200 with the prompt upsampler switched off: 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4. The card also says not to use the model without its prompt upsampler (flux-3-image-klein-pup, a 9B model and an 18.8 GB download), because short prompts give generic results. It says the upsampler adds about 28 GB on an 80 GB GPU. It also warns that file size does not tell you how much GPU memory a variant will use.
These figures were measured under different conditions (a quantised build in ComfyUI versus a reference pipeline on a data-centre GPU), so they don’t cancel each other out. But if you plan to run the recommended pipeline, upsampler included, on one 8–16 GB card, test it before you promise anyone a local setup. Running the upsampler separately, or writing longer prompts yourself, are the obvious things to try.
Speed, by BFL’s measurements
BFL published these timings for a 1-megapixel image:
| Setup | Time |
|---|---|
| Distilled + INT8, RTX 5090 (ComfyUI) | about 4 s |
| Distilled + INT8, RTX 4070 (ComfyUI) | under 9 s |
| Distilled, NVFP4 on GB300, 16 steps | 0.55 s |
| Base, 50 steps, GB300 | 1.7 s |
| Base, 50 steps, DGX Spark | 43.8 s |
BFL says serving LoRAs adds about 5% to generation time. It also compares API latency against other providers. Its own footnote explains that Klein was timed on BFL’s servers, while competitors were timed from the client, including polling and network time. That favours Klein, so treat the latency chart as indicative, not like-for-like.
API pricing
If you would rather not self-host, the distilled model is on BFL’s API as flux-3-image-klein. Its pricing page charges per image by output size, with text-to-image and editing priced the same:
| Resolution | US$ per image |
|---|---|
| 768 × 768 | 0.008 |
| 1K | 0.009 |
| 1.5K | 0.011 |
| 2K | 0.014 |
| 4K | 0.036 |
BFL bills in credits at 1 credit = US$0.01. The API version includes prompt upsampling and “pixel restoration”, which BFL says keeps unedited regions exactly as they were.
How good is it?
All the quality figures so far come from BFL’s internal evaluations. BFL says Klein ranks first in text-to-image, multi-reference generation and editing, ahead of larger open-weight models, and rates above the closed Ideogram V4.5 and FLUX.2 [pro]. It also reports gains over its own FLUX.2 [klein] 9B of +91 Elo in text-to-image and +70 in editing, with 38% fewer parameters. The open models it compares against are Qwen-Image 2.1 (7B) and Krea 2 (12.9B), at the parameter counts their publishers report.
These are vendor-run tests, so run your own prompts, especially for Indian-language text on packaging, signage or ads. The model card itself says text rendering may be inaccurate. It also says bounding boxes guide placement but are not hard masks, so an edit can change content outside the box.
Getting started
- ComfyUI and Hugging Face Diffusers support it from day one. For Diffusers, the card says to install from the GitHub main branch, and its example uses
ModularPipeline. - BFL’s reference code is on GitHub, including a Gradio demo for bounding-box control. Any inference stack you use must support flow-map sampling.
- For LoRA training, BFL points to Ostris’s AI Toolkit and NVIDIA NeMo recipes, and lists TensorRT-LLM, SGLang Diffusion and vLLM-Omni for serving.
- The reference CLI always loads an NSFW image classifier. Keep a safety filter in any pipeline you expose to users.
Who should try it
Teams under the revenue line that generate a lot of product shots, ad variants or edits, and that have a GPU they have tested the full pipeline on (local or cloud), have the most to gain: they get a self-hosted model they can fine-tune with LoRAs. Larger companies, or anyone who needs the upsampler-quality pipeline without managing GPUs, should compare the per-image API price with what they pay today.
References
- FLUX 3 Image Klein: open weights, 5.6B parameters — Black Forest Labs, accessed 2026-10-17
- black-forest-labs/flux-3-image-klein (model card) — Hugging Face / Black Forest Labs, accessed 2026-10-17
- BFL API pricing — Black Forest Labs Docs, accessed 2026-10-17
This week in AI, every Sunday
One email with the releases that mattered and what they mean in practice. No spam.