FLUX 3 Image Klein vs FLUX 3 Image: One-Fifth the Price, and the Comparison BFL Didn't Publish

FLUX 3 Image Klein and FLUX 3 Image are two Black Forest Labs image models with an identical feature list and a five-fold price gap. FLUX 3 Image has been callable at api.bfl.ai/v1/flux-3-image/ since October 1, 2026 at $0.048 for a 1K image. FLUX 3 Image Klein arrived on October 15, 2026 at $0.009 for the same resolution — with its weights published for download, which the larger model's are not. Both accept a prompt, up to ten reference images, and bounding boxes on a 0–1000 grid. Both do native 4K.

So the interesting question is not which one is better, because BFL has not published a head-to-head that would settle it. The interesting question is which one you should be paying for, and the answer depends on a number BFL's launch post conspicuously omits.

What actually landed on 15 October

The weights appeared first. Four Hugging Face repositories went up under black-forest-labs/ before there was anything to read about them: flux-3-image-klein-base/ and flux-3-image-klein/ on October 13, flux-3-image-klein-comfy/ on October 14, and the prompt upsampler flux-3-image-klein-pup/ later the same day. All four are gated behind a contact-information agreement, and at the time of writing none of them carried a public announcement.

That changed on October 15, when BFL published FLUX 3 Image Klein: open weights, 5.6B parameters and opened the API endpoint. The reference inference repository on GitHub was created the same afternoon. So the radar's read — a checkpoint that existed before the launch post — was correct about the two-day head start and then overtaken by events: this is an announced release, published yesterday, not a leak that the vendor is still sitting on.

What shipped, concretely:

• Two checkpoints at 5.6B parameters each. Klein Base is undistilled, runs 50 steps at guidance 3.5, and exists for fine-tuning, LoRA training and research. Klein is step-distilled to 16 steps with no guidance and is the one behind the API. "5.6B" counts the image transformer only; the text encoder is Qwen3-VL-4B-Instruct and is a separate component.

• Quantised builds. FP8 and NVFP4 variants ship in the repository alongside the BF16 transformer, with INT8 built with NVIDIA support. The generator needs 11.2 GB in BF16 and 5.6 GB in FP8.

• A commercial licence with a revenue ceiling. The FLUX Kommunity License permits free commercial use for individuals and companies under $5M in annual revenue. Above that, self-hosted deployment goes through BFL's paid licensing. This is a real restriction, not a formality, and it is the single biggest practical difference between downloading Klein and renting FLUX 3 Image.

Where the two are genuinely identical

The feature parity is not marketing compression — BFL's own documentation states that Klein "supports the same features" as FLUX 3 Image, and lays them out as the same four capabilities. Reading the two prompt surfaces side by side:

• Prompt interface — prompt/ is required, everything else optional, on both.

• References — up to 10 images, URL or base64, accepted on both. FLUX 3 Image takes each reference between 256×256 and 16 megapixels; references above roughly 4 MP are downscaled. Klein is documented as taking up to 10 references with the same multi-reference composition behaviour.

• Bounding boxes — supported on both, on the same 0–1000 grid, for placing elements and for confining edits box by box. BFL is explicit that boxes guide placement and scale rather than constraining pixels to a hard mask.

• Maximum resolution — 4K on both. The resolution ladder is the same five rungs: 768×768, 1K, 1.5K, 2K, 4K.

• Editing — instruction-based edits that preserve unedited regions, and multi-turn box-by-box edits, on both.

If your integration only touches those five things, the two models are drop-in replacements for each other and you are choosing purely on price, latency and licence.

Where they diverge, and it is not quality

• Weight availability — Klein only. FLUX 3 Image is API-only. There is no downloadable FLUX 3 Image checkpoint, and BFL's own licensing page sells self-host tiers for the FLUX.2 family rather than for 3. Once you can download Klein, "the model is slow" and "the model is unavailable" stop being the same complaint.

• Price per image — 5.1× to 16.9× apart, widening with resolution. At 768sq it is $0.008 against $0.041. At 1K, the tier most integrations default to, it is $0.009 against $0.048. At 2K it is $0.014 against $0.10, and at 4K it is $0.036 against $0.607 — a factor of nearly seventeen.

• Hardware floor — Klein only. With the INT8 build, Klein runs on consumer RTX cards with as little as 8 GB of VRAM, and the docs put the floor at an RTX 3090.

• Licence — different instruments. FLUX 3 Image is a metered commercial service with a published per-image rate and a 30-day retention-shaped API contract. Klein is a weights licence with a $5M revenue threshold.

The 4K gap is the one worth staring at. BFL prices FLUX 3 Image at $0.607 per 4K image and Klein at $0.036 for the same resolution. If you are generating 4K at any volume, that difference is not a discount, it is the difference between a viable product and an unviable one — and it is available on the smaller model.

The comparison BFL published does not contain FLUX 3 Image

This is the part worth reading the launch post closely for. BFL's own summary of Klein's quality reads: it "ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models," and "also rates above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro]."

The comparison set is named elsewhere in the post: Qwen-Image 2.1 at 7B, Krea 2 at 12.9B, FLUX.2 [klein] 9B, plus Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B, Klein gains 91 Elo in text-to-image and 70 in image editing while carrying 38% fewer parameters. Krea 2 does not support image editing and sits out that column.

FLUX 3 Image — the vendor's own larger, more expensive model, on the same feature surface — appears nowhere in that list. That is not necessarily damning. It may simply be that BFL benchmarked Klein against the models a Klein buyer would otherwise download. But it does mean that the sentence a buyer most wants, "Klein is within X of FLUX 3 Image," has not been written by anyone, and the price ladder is being sold on a quality claim that stops short of the tier above it.

Treat every figure here as vendor-reported. These are BFL's internal evaluations, run by BFL, published by BFL. No independent leaderboard has scored either model yet. The one published comparison that exists at all is theirs.

The speed numbers come from NVIDIA, and the comparison is not apples to apples

Klein's headline latency figure: with the distilled model and the INT8 build, about 4 seconds for a 1 MP image on an RTX 5090 and under 9 seconds on an RTX 4070, through ComfyUI. On GB300 in NVFP4, BFL quotes under 0.6 seconds for the same 1 MP output. Serving a LoRA adds roughly 5% to generation time. These are measured by NVIDIA, not by BFL, and not by us.

The API latency comparison deserves more scepticism than it usually gets, because BFL flags the problem in their own footnote. Klein was timed on BFL's servers. The models it is compared against — Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5, Krea 2 — were timed from the client, which includes polling and network time. The blog states this plainly. It means the frontier chart showing Klein fastest at its quality level is measuring two different things on the two sides of the line.

Krea 2's own row illustrates the spread you get from configuration alone: BFL timed it on its Medium tier, and notes that its Large tier took 31.1 seconds for text-to-image.

Which one to call

Pick FLUX 3 Image when you need the largest model's output and you are paying per image at low volume, or when you cannot accept a licence with a revenue ceiling, or when 4K volume is negligible. It is the safer default if you have no GPU and no appetite for the Kommunity License's terms.

Pick FLUX 3 Image Klein when your per-image cost matters more than the top of the quality range — which at 4K it almost always will, given the 17× spread — or when you need the model to run somewhere BFL's API cannot reach: on-premises, air-gapped, or on a laptop with 8 GB of VRAM.

There is a third option worth naming, because a one-day-old open-weights release is exactly the case it exists for. If you are already calling other image models through an OpenAI-compatible endpoint, that endpoint is where failover and price pass-through live: a provider repricing shows up without a redeploy, and you can try an unfamiliar model on a fraction of traffic without rebuilding your client. OrcaRouter routes a couple of hundred models — including the image tier of google/gemini-3.1-flash-image-preview/ — at provider list price with nothing added on top. FLUX 3 Image Klein is not one of them: BFL's own API is the only route to Klein's hosted endpoint, and the weights are the other. We are not going to pretend otherwise.

What is still unknown

Klein is a day old and the honest list of gaps is long.

• No independent benchmark. Nothing on Artificial Analysis or any other neutral board scores either model. Every quality claim in this article is BFL's.

• No Klein-versus-FLUX-3-Image number. Not from the vendor, not from anyone else.

• No long-run uptime or throughput data. One day of API traffic tells you nothing about either.

• Adoption is unmeasured. The distilled repository showed 3,241 downloads in the last month at the time of capture, on a repo that is gated. That number is small, and it is not a quality signal.

• The safety evaluation is a vendor claim. BFL states the checkpoints were red-teamed internally and evaluated by an external partner, Cinder, and reports "substantially fewer vulnerabilities than other popular open weights models" — without publishing a number.

What would settle most of the above is a third party running both models on the same prompts at the same resolutions. Until that exists, the price gap is the only hard, verifiable comparison in this article, and it is the one to buy on.

The thing to watch next is not Klein's benchmark score. It is whether BFL positions it as the value tier or lets it compete with its own flagship from below. A 5.6B open model at one-fifth the price with the same feature list is a strange thing to ship next to a $0.607 4K tier unless the intent is that buyers migrate. If FLUX 3 Image quietly acquires a cheaper tier, or if the Klein comparison chart ever gains the flagship as a row, that will be the tell.