FLUX 3 Image Klein Runs on an 8 GB GPU. Its Required Prompt Upsampler Needs About 28 GB More.
Black Forest Labs released FLUX 3 Image Klein on October 15: a 5.6B open-weight image model with native 4K, up to ten reference images, and an API price of $0.009 per image at 1K. BFL says INT8 runs on 8 GB RTX cards, and BFL's own docs say not to use the model without a separate prompt upsampler that adds about 28 GB of GPU memory. No independent quality score exists yet, and the license caps commercial use at $5 million in revenue, requires content filters, and can be revoked on notice. We read the blog, the docs, the Hugging Face repos and license, AA's board, and the Gateway prices.
By T.W. Ghost
Black Forest Labs released FLUX 3 Image Klein today, October 15, 2026: a 5.6-billion-parameter image model with open weights, native 4K output, up to ten reference images, bounding-box control, and an API price of $0.009 per image at 1K resolution. The pitch is consumer hardware. BFL says the INT8 build runs on RTX cards with 8 GB of VRAM, and quotes NVIDIA's measurement of about four seconds per megapixel image on an RTX 5090.
Then BFL's own documentation adds a sentence the blog does not: "Do not use it without its prompt upsampler." The upsampler is a separate model, about 9.4B parameters based on Qwen3.5, and the docs say it adds about 28 GB of GPU memory. The same docs say the model fits consumer GPUs "down to an RTX 3090." Eight gigabytes is the diffusion model alone. We read the blog, the docs, the three Hugging Face repos and the license file, Artificial Analysis's text-to-image board, and Vercel's Gateway price rows. Here is what held up.
What You Are Downloading
| FLUX 3 Image Klein | |
|---|---|
| Released | October 15, 2026 |
| Size | 5.6B parameters; base and step-distilled variants; FP8, INT8, and NVFP4 builds; 11.2 GB BF16, 5.6 GB FP8 |
| Features | Native 4K; up to 10 reference images; bounding-box control |
| Hardware (BFL) | INT8 on 8 GB RTX cards; about 4 s per 1MP image on an RTX 5090, under 9 s on an RTX 4070 (distilled, INT8, NVIDIA's measurement); under 0.6 s on GB300 with NVFP4 |
| Required companion | Prompt upsampler black-forest-labs/flux-3-image-klein-pup, about 9.4B parameters, Apache 2.0, ungated; adds about 28 GB of GPU memory per BFL's docs |
| Repos | flux-3-image-klein and -klein-base gated with auto-approval, license "other" |
| License | FLUX Kommunity License (see below) |
| Red team | Cinder (BFL's statement) |
| AA text-to-image board | No Klein row as of today |
The hardware claim has two halves. The 8 GB figure describes the diffusion model. BFL's docs tell you not to run it without the upsampler and give the upsampler's memory cost at about 28 GB, then describe the whole thing as fitting "down to an RTX 3090." Whether the 8 GB line assumes the upsampler runs on another machine is not stated. The upsampler is Apache 2.0 and separate, so you can host it elsewhere; the point is that you have to host it.
The Price Table
| Resolution (BFL API, per image) | Price |
|---|---|
| 768 | $0.008 |
| 1K | $0.009 |
| 1.5K | $0.011 |
| 2K | $0.014 |
| 4K | $0.036 |
Vercel's AI Gateway lists bfl/flux-3-image-klein at the same per-size prices. For scale, full FLUX 3 Image on the Gateway is $0.048 at 1024 and $0.607 at 4096, so Klein at 4K is about 94% cheaper than the full model at 4K, by our arithmetic. The blog's "$0.009 per 1K image" means per image at 1K resolution, and 4K costs four times that.
The Quality Claim Is BFL's
BFL says Klein "ranks first" in its internal evaluations against its own comparison set, with +91 Elo on text-to-image and +70 on editing over FLUX.2 Klein 9B. Those are BFL's numbers.
Artificial Analysis has not listed Klein. For context on its board today: full FLUX 3 Image sits 12th at 1,116 Elo (plus or minus 13); the top open-weight model is Qwen-Image-2.1 at 21st with 1,042; FLUX.2 Klein 9B is 72nd at 931. Where Klein lands between its predecessor and the full model is the number this post cannot print, and the one to re-check before you commit to the model.
The License, Read From the File
The FLUX Kommunity License on the base repo:
- Commercial use of outputs only for a "Qualifying User," defined as you and your affiliates with under US$5,000,000 in gross annualized revenue "at any time."
- Qualifying Users must implement content filtering.
- Outputs may not be used to improve competing models; distillation into other models counts as a derivative.
- BFL "may terminate this License, in whole or in part, at any time upon notice."
Above $5 million, by the grant's own terms, you are on non-commercial use unless you sign a separate deal. "Open weights" here means gated download, a revenue cap that counts affiliates, a filtering requirement, and a license BFL can end.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 15; 5.6B; native 4K; 10 reference images; FP8, INT8, NVFP4 builds | Verified (BFL blog and docs) |
| INT8 runs on 8 GB RTX cards; about 4 s per 1MP on an RTX 5090 | Vendor (BFL, quoting NVIDIA) |
| "Do not use it without its prompt upsampler," which adds about 28 GB | Verified (BFL docs, quoted) |
| Fits "down to an RTX 3090" | Verified as BFL's docs wording; conflicts with the 8 GB line |
| API prices by resolution; $0.036 at 4K; Gateway matches | Verified (BFL docs; Gateway API) |
| "Ranks first," +91 and +70 Elo | Vendor (BFL internal evaluations) |
| Independent quality score | Unverified (no AA row yet) |
| $5M revenue cap incl. affiliates; content filters; revocable on notice | Verified (LICENSE.md on Hugging Face) |
Who Should Use It
Use Klein on the API as the cheap default for image work: $0.009 at 1K and $0.036 at 4K are hard to beat, and you can test it against full FLUX 3 Image on your own prompts without a VRAM budget. Use it locally if you can host both models, or run the Apache 2.0 upsampler somewhere else and the 8 GB diffusion model on the card BFL advertised.
Skip it if you are at or near $5 million in revenue across affiliates and have not talked to BFL, if you cannot implement the required content filtering, or if you need an independent quality number before you ship. And re-check AA's board; a Klein row is the missing fact.
The frontier is elsewhere: on AA's text-to-image board the top is closed models, full FLUX 3 Image is 12th, and the best open model is 21st. Klein's place in that order is not known today. Our model comparison has the current standings side by side.
Sources
- Black Forest Labs: FLUX 3 Image Klein
- BFL docs: FLUX 3 Image Klein
- Hugging Face: black-forest-labs/flux-3-image-klein
- Hugging Face: black-forest-labs/flux-3-image-klein-base (license)
- Hugging Face: black-forest-labs/flux-3-image-klein-pup
- Artificial Analysis: text-to-image leaderboard
- Vercel AI Gateway: models
*Not sure whether a cheap local image model or a hosted one fits your pipeline? Take the free 2-minute quiz and get matched. Then read about the video model from the same family that fal shipped four days later, with a price cliff of its own: FLUX 3 Max.*