The 5.6B label counts part of the system
Klein is the small member of the FLUX 3 Image family, and small is relative to a specific component. The model card states plainly that the 5.6B figure refers to the image transformer. The text encoder that reads the prompt is a separate model, Qwen3-VL-4B-Instruct. The reference command-line setup also loads a safety classifier every time, and its optional prompt upsampler, the component that rewrites a short request into the full caption the model expects, is an 18.8GB download by itself. That 18.8GB is a download size, not a memory requirement: the card measures the upsampler adding about 28GB of GPU memory on an 80GB GPU when it actually runs.
None of that makes Klein heavy by current standards. It does mean the parameter count on the launch graphic is not the size of the thing you install, run and keep in memory. Anyone sizing a machine, or comparing Klein against a rival quoted as a single total, should compare full stacks.
8GB is a floor with nothing attached to it
The consumer-hardware pitch, in BFL's own terms, is that an INT8 build runs through ComfyUI or Diffusers on GeForce RTX cards with as little as 8GB of VRAM. Black Forest Labs has not tied that minimum to a checkpoint, an output size or a generation time, and its supplied speed charts do not test an 8GB configuration. Read it as a minimum supported deployment, not a measured one.
Its own published numbers show how wide the range is, and they are BFL-reported, not independently tested. On an RTX 4070 using INT8, the distilled model takes 8.9 seconds to generate a 1024×1024 image and 18.9 seconds to edit one. At 2048×2048 those figures become 50.4 seconds and 150 seconds. Separately, the model card's reference implementation, measured on a B200 at 1024×1024 with the distilled checkpoint and the prompt upsampler switched off, peaks at 31GB in BF16, 26GB in FP8 and 25GB in NVFP4. Those peaks are a different measurement from the 8GB floor, not a contradiction of it — but they are the only memory numbers BFL actually measured in public. A quantized ComfyUI setup with offloading may well fit into 8GB. What it produces there, and how long it takes, is the part a buyer still has to test.
The quality claims need the same patience. The first-place Elo ratings come from BFL's internal testing: +91 Elo points for text-to-image and +70 for editing, both against FLUX.2 [klein], the 9B predecessor, against a comparison set that stops short of several current rivals, with no prompts, judging method or sample sizes published alongside. The charts are a reason to try the model. They are not a ranking anyone else has verified.
Open weights, with a revenue line drawn through them
The licence is where Klein separates itself from both closed APIs and genuinely permissive releases. Under the FLUX Kommunity License, use is non-commercial by default. The exception covers commercial output generation by qualifying users whose gross annualized revenue, affiliates included, is under US$5 million. Output generation is the operative phrase. The model card adds that this exception is not blanket permission for commercial fine-tuning or distribution, and the licence's out-of-scope list includes using the model or its outputs to improve a competing image model.
The boundary is a real one because BFL sells the other side of it: paid commercial licences, including self-hosted licences that explicitly include fine-tuning rights, cover the uses the free exception leaves out. The model card points anyone outside the exception to BFL's commercial licensing.
For an independent creator or a small studio, the output side is still a workable deal: download the weights, generate locally, sell the pictures. Training a LoRA to sell is a separate question — qualifying to generate commercial outputs does not by itself qualify anyone for commercial fine-tuning, and the card says so plainly. For a company anywhere near the threshold, the affiliates clause means a small team's project does not reset the calculation. The revenue of the group is what counts.
One price shrank a thousandfold in the retelling
A widely carried launch report, The Decoder, priced the API at $0.009 per 1,000 images. BFL's own API documentation prices Klein per image: $0.009 buys one 1K image, which makes 1,000 images at 1K cost $9, not $0.009. The full list runs $0.008 per image at 768×768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K and $0.036 at 4K, with generation and editing priced the same at each size. The real price is competitive against hosted rivals at the same sizes. It is also a thousand times the figure that travelled. Any cost model built this week on the per-thousand version needs redoing, with prices as checked on 15 October 2026.
What you download is the model, not the pipeline
What Klein gets right is the shape of the release — but read the card's own description of what the download is. In BFL's words, the files "contain the model package, not a complete inference application": the distilled and base checkpoints in BF16, FP8 and NVFP4, the text encoder and the image autoencoder. The card also notes the distilled checkpoint is "the checkpoint behind the BFL API". Same weights; different thing. The API is BFL's production pipeline around those weights, where BFL runs the serving stack, the prompt upsampling and the safety filtering. Self-hosting means assembling that pipeline yourself, starting with an inference implementation that supports flow-map sampling and this checkpoint format, and deciding which of the surrounding pieces — upsampler, safety classifier — you run and how.
Access has a gate as well. The Hugging Face repository is publicly listed, but the files sit behind the platform's gated-access step: you accept the licence conditions and share contact details before anything downloads. Open weights, here, does not mean an anonymous one-click download.
The release still prints its limitations on the card: bounding boxes guide placement rather than masking it, and edits can change content outside the region you asked about. Test it on your own images, at the size you actually ship, under the licence tier you actually fall into.
Sources
- [1] Hugging Face model card, black-forest-labs/flux-3-image-klein (primary source — 5.6B transformer scope, separate Qwen3-VL-4B-Instruct text encoder, checkpoint table naming the distilled build "the checkpoint behind the BFL API" and the download as "the model package, not a complete inference application", reference hardware peaks of 31GB BF16 / 26GB FP8 / 25GB NVFP4 at 1024×1024 with upsampler off, 18.8GB prompt upsampler download adding about 28GB of GPU memory, gated file access, bounding-box and edit limitations, FLUX Kommunity License v1.0 terms and the $5M qualifying-user threshold)Read source
- [2] Hugging Face model card, black-forest-labs/flux-3-image-klein-comfy (primary source — ComfyUI repack contents: distilled and base transformers in BF16, INT8 and NVFP4)Read source
- [3] VentureBeat, "Black Forest Labs opens FLUX 3 Image Klein weights, with free commercial use limited to smaller businesses" (15 October 2026, 15:50 UTC — BFL API price list by resolution, RTX 5090 and RTX 4070 INT8 timings from BFL charts, the unattributed 8GB minimum, benchmark chart limits, and the $5M licence reading)Read source
- [4] Black Forest Labs press release via Globe Newswire, "Black Forest Labs Expands Access to Frontier Visual Intelligence With Open Weights Release of FLUX 3 Image Klein" (15 October 2026 — company announcement, FLUX 3 family context, base / distilled / quantized release structure)Read source
- [5] The Decoder, "Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM" (15 October 2026 — BFL's internal Elo claims, +91 points for text-to-image and +70 for editing over FLUX.2 [klein], under-9-second RTX 4070 claim; this is also the report that carried the $0.009 per 1,000 images figure, contradicted by BFL's per-image API documentation above)Read source