FLUX 3 Image Klein vs FLUX 3: Which Half of the Family You Can Actually Download

You can download FLUX 3 Image Klein. You cannot download FLUX 3 — not the whole thing, and not most of it. The FLUX 3 family that Black Forest Labs announced on July 23, 2026 spans image, video, audio and action prediction in one Self-Flow architecture. Twelve weeks of shipping later, exactly two of its capabilities exist as weights you can pull off Hugging Face: the 7B action head from September 22, and now a 5.6B image model published on October 15. The image tier the family is named for, and the video tier that has been generally available since August 4, remain API-only. The list price you see is the provider's, not ours — zero-markup pass-through which matters the week a vendor reprices.

That split is the real story here, and it is a more useful one than a spec sheet. Klein is not "a smaller FLUX 3." It is one part of FLUX 3, extracted, quantised, and handed over in a form the rest of the family does not come in.

What "FLUX 3" refers to, and why the ambiguity matters

BFL uses FLUX 3 for three different things in the same week's worth of documentation: the research programme described in the July paper, the umbrella over four separately-priced products, and — in several places on the model pages — the specific image model that has its own name, FLUX 3 Image.

The July announcement is the one that explains the architecture. FLUX 3 is trained jointly on images, video and audio because, in BFL's framing, no single modality gives a complete picture of the world: images carry spatial structure, video carries motion and physics, audio carries causal relationships that vision alone misses. The training was not evenly weighted. More than 95% of the compute went to video, and under 0.5% of tokens were allocated to audio — a distribution that tells you what BFL believed the hard problem was, and which is worth holding onto when you read quality claims about any member of the family.

So when Klein's launch post says the model "draws on FLUX 3's shared learning across images, video, and audio," that is a claim about pretraining lineage, not about capability. Klein is an image model. It generates and edits stills. The video and audio training is upstream of it, not inside it.

The four parts, and which two have weights

• FLUX 3 Video — generally available since August 4, 2026, metered per second of output. Clips up to 20 seconds with native synchronised audio, at HD and FHD, with upscaling to higher resolutions. No weights.

• FLUX 3 Action — a 7B world action model for robot control, published September 22, 2026 with the weights and a training report. This was the family's first open-weights release.

• FLUX 3 Image — live at api.bfl.ai/v1/flux-3-image/ since October 1, 2026. Native 2K and 4K, bounding boxes on a 0–1000 grid, up to ten references. No weights, and BFL's licensing page sells self-host tiers for the FLUX.2 generation rather than for this one.

• FLUX 3 Image Klein — published October 15, 2026. 5.6B parameters for the image transformer, distilled and base checkpoints, FP8 and NVFP4 variants, BF16 and INT8 builds produced with NVIDIA support, under the FLUX Kommunity License. The distilled checkpoint is what serves the BFL API endpoint.

Two open, two closed, and no published plan for closing the gap on the other two. The launch post's closing line is "More will follow" — which is a statement of intent, not a schedule.

Open weights change the cost question, not just the cost

Comparing the two on price is awkward because they are not sold in the same unit. FLUX 3 Video bills per second; FLUX 3 Image and Klein bill per image; FLUX 3 Action bills per prediction. Very little of the family can be compared line to line.

What can be compared is the shape of the commitment:

• Rented capability — a per-unit rate, a published price list, no hardware, and a licence you accept by calling the API. FLUX 3 Image runs $0.041 to $0.607 per image depending on resolution; FLUX 3 Video runs $0.06 per second for draft-effort HD and $0.15 per second at standard-effort HD, as published on the BFL pricing page on October 16, 2026.

• Owned capability — a download, a GPU, and a licence with a revenue threshold. Klein's generator needs 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 build it runs on consumer RTX cards from 8 GB of VRAM, and the documentation puts the floor at an RTX 3090.

The Kommunity License is the part teams under-read. Free commercial use covers individuals and companies with annual revenue below $5M; above that, self-hosted commercial deployment goes through BFL's paid licensing. A startup that downloads Klein today at $2M revenue and grows past the threshold inherits a licensing conversation it did not plan for. That is a different kind of risk from a per-image rate, and it is the one that bites later.

For anyone weighing that switch, the practical question is not "which is cheaper" but "which of my workloads can move." There is a real middle path: keep the metered calls on a single endpoint so a price change or an outage does not require a client rewrite, and add anything you self-host behind the same interface. That is what routing is for — one API key across a few hundred models with adaptive routing and failover — and it is how you evaluate a one-day-old checkpoint on five percent of traffic instead of committing a production path to it. Klein's hosted endpoint is not on OrcaRouter; we route the image tier of google/gemini-3.1-flash-image-preview/ and similar models, and BFL's API plus the weights are the routes to Klein itself.

What Klein inherits that the API tier does not need to care about

The most consequential thing about Klein shipping as weights is not the price. It is that the constraints stop being BFL's problem and become yours.

• Prompt upsampling is a separate download. The reference implementation loads a prompt upsampler — flux-3-image-klein-pup/, Apache-2.0, roughly 18.8 GB — that rewrites your prompt before generation. Skip it (--no-pup/) and the model takes prompts as typed. Through the API, BFL makes that decision for you.

• A safety classifier is always loaded. The reference CLI pulls a separate NSFW image-detection model and loads it unconditionally. On the API side, moderation is a line item with defined billing behaviour; on your own hardware it is a component you now maintain.

• Quantisation is your choice with quality consequences. FP8, NVFP4 and INT8 are shipping in the repository, and the manufacturer's own note says a quantised package's file size is not a measure of total GPU memory use. Picking a build is a quality-versus-VRAM trade that the API does not expose at all.

• Sampler settings are documented, not enforced. Distilled runs 16 steps with no guidance; Base runs 50 at guidance 3.5. The docs also give the settings for running a Base-trained LoRA on Distilled — 16 steps, guidance 1, LoRA scale around 0.75 — and note that LoRAs transfer between the two without conversion. None of that has an API equivalent.

This is the honest case for the API tier even at 5× the price: it is a fully assembled product, and the weights are a kit.

The gap nobody has measured

Klein's quality claim, as BFL states it, is that internal evaluations put it first in text-to-image, multi-reference generation and image editing, ahead of larger open-weights models, and above the closed models in the comparison — 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.

Read the comparison set carefully and you notice what is missing: no FLUX 3 Image, no FLUX 3 Video, nothing from the family Klein belongs to. The evaluation is against other open models and two third-party closed ones. So the question a family-level comparison invites — how does a 5.6B downloadable model close against the API-only tier it shares a name with — is unanswered by the vendor and unanswered by anyone else.

There is one more thing the launch post does not publish, and it is cited as an external check: BFL says the Klein checkpoints were red-teamed internally and evaluated by a partner, Cinder, showing "substantially fewer vulnerabilities than other popular open weights models." No figure accompanies that. An external evaluation is worth more than an internal one; an external evaluation without a number is worth less than it looks.

Everything above is vendor-reported. Neither FLUX 3 nor FLUX 3 Image Klein appears on an independent leaderboard as of October 16, and if you need a neutral number before committing, none exists yet.

How to decide, and what would change the answer

Take the downloadable model if your constraint is deployment location rather than quality — anywhere BFL's API cannot reach, from a 8 GB laptop GPU to an air-gapped cluster — or if your monthly image volume at 2K and above would make the metered rate the dominant line in your budget. Klein at $0.036 per 4K image against FLUX 3 Image at $0.607 is a seventeen-fold difference; if 4K is your default, that arithmetic answers the question before quality enters it.

Stay on the API if you want the largest model BFL will sell you, if you would rather not run a safety classifier and a prompt upsampler as your own services, or if your revenue trajectory makes a $5M licence ceiling an uncomfortable thing to have in your architecture.

Three things would settle the family question cleanly, and none of them exist yet: a third-party run of Klein against FLUX 3 Image on identical prompts; a BFL statement of intent about whether FLUX 3 Image and Video weights will ever ship; or a published Cinder number. Until then the only firm facts are the ones you can check yourself — the repositories, the price list, and the fact that half of FLUX 3 is still something you rent.

The near-term tell is whether a third open-weights release arrives on the same cadence. Action came in September, Klein in October. If November brings FLUX 3 Video weights, the family stops being a set of products and starts being a platform, and every argument in this article about renting versus owning changes shape. If nothing follows, Klein is a one-off in the image line and the closed tiers are the strategy.