FLUX 3 Klein Says 8 GB of VRAM. The File Sizes Say Otherwise.

Black Forest Labs released FLUX 3 Image Klein on October 15, and the headline number is 8. As in eight gigabytes of VRAM, supposedly enough to run a 5.6-billion-parameter image model with native 4K output, ten reference images, and bounding-box layout control. For anyone who has spent the last two years watching open image models require a 24 GB card and a prayer, that reads like the end of a long wait.

The number is also, in the way these numbers usually are, technically true and practically misleading. Both halves of that sentence deserve attention, because Klein is genuinely interesting and the gap between its marketing and its file sizes tells you a lot about where local image generation actually stands.

What shipped

Klein is the open-weights member of the FLUX 3 family. The FLUX 3 Image model that arrived on October 1 was API-only; Klein is the one you can download. Its architecture is an image transformer at 5.6 billion parameters, paired with a Qwen3-VL 4B text encoder that carries an Apache 2.0 license — a detail worth noting, since the main weights are not Apache.

Three capabilities come in one set of weights: text-to-image, single-reference editing, and multi-reference generation with up to ten images in a prompt. The reference limit is the headline for production work. If you're trying to keep the same actor, the same costume, and the same prop consistent across a sequence of frames, ten slots is enough to assign one job per reference instead of hoping a mood board averages out correctly.

The layout control is bounding boxes on a 0-to-1000 grid. You draw a region, and tell the model to place, recolor, move, replace, or remove whatever is inside it. This is a different interaction from prompt-and-pray. It's closer to a page layout tool, which is why the release reads less like a model launch and more like a feature addition to a design workflow.

BFL also ships two checkpoints. A distilled one, 16 steps with no guidance, intended for inference. A base one, 50 steps at CFG 3.5, intended for training — the model card is explicit that the distilled checkpoint is a bad starting point for fine-tuning. The documented LoRA recipe trains on Base and serves on Distilled, and because the two share an architecture, the adapter loads without conversion. In BFL's worked example, a style LoRA on twenty captioned public-domain paintings at rank 32 took about 4.5 hours on one GPU.

The 8 GB claim, itemized

Here's the arithmetic the announcement doesn't do for you. The ComfyUI repository lists the component files:

The transformer is 11.1 GB in BF16, or 5.8 GB in INT8. The Qwen3-VL text encoder is 8.9 GB BF16, 4.9 GB INT8. The VAE is 0.3 GB. And then there's the prompt upsampler — a Qwen3.5-9B-based model at 18.8 GB BF16 or 9.8 GB INT8.

Take the two biggest INT8 pieces: 5.8 plus 4.9 is 10.7 GB of weights before a single activation is allocated. An 8 GB card cannot hold both. So the 8 GB claim depends on offloading one of them to system RAM, which works, and which is also why the vendor's own model card measures a 31 GB peak in BF16 and 26 GB in FP8 at 1024×1024 on a B200 — with the upsampler switched off.

That last clause matters more than it looks. The model card states plainly that a short prompt typed as-is produces generic output, and that the upsampler is required for good results. The upsampler needs roughly 28 GB at full precision on an 80 GB card. This is not an optional luxury component. It's part of how the model is meant to be used, and it's roughly the size of the model itself.

So the honest summary is: 8 GB gets you a run, on an INT8 build, with aggressive offloading, at reduced speed, and probably with a weaker prompt path. It does not get you the experience in the launch video.

What the speed numbers do and don't say

BFL reports about four seconds for a one-megapixel image on an RTX 5090, and under nine seconds on an RTX 4070. Those are good numbers. They're also measured on the distilled checkpoint, with the settings BFL chose, for a single image, excluding model load. If you're running a workflow that swaps between models — an upscaler, an inpaint pass, a face restoration step — the wall-clock time per finished asset will be several times that.

The vendor also claims Klein beats several larger open-weight models on text-to-image, multi-reference, and editing benchmarks. That claim is unverified by third parties as of this writing, and internal comparisons at launch are the least reliable genre of benchmark. The independent leaderboard numbers that exist put Klein mid-pack on raw aesthetics and notably better on edit adherence, which fits the design priorities.

The license is the part people skip

Klein's weights are under the FLUX Kommunity License, not Apache 2.0. Commercial output rights are granted to individuals and companies whose gross annualized revenue, counted with affiliates, stays under 5 million US dollars. Above that, you need a commercial license from BFL or you use the API.

For a solo designer, a small studio, or a filmmaker cutting a music video, that's a real grant. For an agency that just crossed the threshold, it's a compliance item that shows up in an audit two years later. "Open weights" and "open source" are being used interchangeably in most of the coverage of this release, and they are not the same thing.

There's also a pricing inversion worth doing the math on. The API starts around 0.008 to 0.009 dollars per image. At a thousand images a month, that's under ten dollars. A 24 GB card costs more than a thousand. Self-hosting Klein only makes financial sense at volumes where the API bill is the bigger number, or in situations where the images can't leave your machine.

Who should actually download it

If your work needs character or product consistency across a set of images — storyboards, a product catalog with one object in twenty scenes, a comic where the same face appears in every panel — Klein's ten references plus box control is the most direct tool available in open weights right now. That combination is genuinely new at this size.

If your work needs one beautiful hero image, the API is cheaper, faster, and better than your local setup will be.

If your reason for wanting local weights is privacy — medical, legal, unreleased product — then the 31 GB BF16 peak is the number to plan around, not eight. Budget a 32 GB card, or wait for the community GGUF quantizations that always arrive within a month or two.

The pattern with open image models has been consistent for two years: the announcement quotes the smallest possible footprint, the model card quotes the real one, and the community splits the difference with quantized builds that trade a little quality for a lot of headroom. Klein follows that pattern exactly. Knowing which number applies to you is most of the decision.