FLUX 3 Image Klein Ranks #1 Among Open Weights — on the Scoreboard Its Maker Drew

The number is everywhere now. FLUX 3 Image Klein finished first among open-weight models in text-to-image, first in multi-reference generation, first in image editing — and beat the older FLUX.2 [klein] 9B by 91 Elo in text-to-image and 70 Elo in editing while carrying 38% fewer parameters. Those are the figures in Black Forest Labs' launch post, in the recaps, in the social card. The sentence that tends to get dropped is the one attached to them in the source: these are internal evaluations.

That is not a scandal. It is how almost every image model launches, and it is the only comparison a lab can run before anyone else has the weights. But it changes what a reader should do with the ranking. FLUX 3 Image Klein is genuinely released — 5.6B parameters, open weights, published 15 October 2026 under the FLUX Kommunity License, callable at $0.009 per 1K image on BFL's own API. What does not exist yet is a score from anyone who is not Black Forest Labs. The model is real; the leaderboard position is a claim.

What the release actually established

Separate the artefacts from the marketing, because they carry different weight. The artefacts are checkable today, by anyone, without a GPU:

• Weights — a 5.6B-parameter image generator, with a base checkpoint for fine-tuning and a 16-step distilled checkpoint for latency

• Precision builds — BF16 at 11.2 GB, FP8 at 5.6 GB, plus INT8 and NVFP4 builds made with NVIDIA

• Licence — FLUX Kommunity License: free commercial use for individuals and companies under $5M annual revenue, paid self-hosted licensing or the API above that line

• Capabilities — text-to-image, editing, multi-reference generation with up to 10 reference images, native 4K, and bounding-box control

• Price — $0.009 per 1K image on the vendor API, with a 2K tier above it

• Tooling — LoRA training recipes for Diffusers, Ostris AI Toolkit and NVIDIA NeMo AutoModel

All of that is in the repository and the docs. None of it depends on a benchmark being true. A model can be small, fast, licensable and downloadable without being the best at anything, and those properties are worth having on their own.

What the release did not establish is relative quality. The launch post shows Elo charts rather than printing exact scores in the text, so the precise gap between Klein and its comparison set is not even fully quotable — you can read the bars, not the numbers. The named comparators are Qwen-Image 2.1 at 7B and Krea 2 at 12.9B, with Krea 2 excluded from the editing comparison because it does not support editing. Those are sensible controls. They are still the lab's own controls, in the lab's own harness, on prompts the lab chose.

The word "third-party" is doing work it has not earned

A lot of the coverage of this release describes the rankings as third-party benchmarks. They are not. A search for an independent leaderboard entry turns up the vendor's blog post and a dozen rewrites of it, and the rewrites inherit the framing without inheriting the caveat.

The most useful public board for image models is Artificial Analysis, which ranks by Elo from pairwise human preference votes. Its text-to-image board, version 2.1, captured on 16 October 2026, has 42 visible entries. GPT Image 2.5 Sunburst sits first at 1218. Nano Banana 2 is seventh at 1139. FLUX 3 Image is thirteenth at 1114, on 5,766 samples, priced by the board at $48 per thousand images. FLUX.2 [klein] 9B — the baseline in BFL's comparison — is twenty-sixth at 1026.

FLUX 3 Image Klein is not on that board. Not at launch, not eight days later. The entry labelled "FLUX 3 Image" is a different product: it is the larger API-only sibling, and the board's $48-per-thousand figure matches that model's published range, not Klein's $9. This matters for a simple reason. On the one independent ranking that exists, the highest-scoring open-weight image model is not Klein and is not a Black Forest Labs model at all — it is Qwen-Image-3.0-Pro at 1093. The claim that Klein is first among open weights is therefore not contradicted by the public board. It is simply absent from it.

Put BFL's own delta next to a public board

There is a way to get a rough bearing without pretending it is a measurement. BFL states its gains as deltas against FLUX.2 [klein] 9B, and that older model is on the public board. So take the public score and add the vendor's delta:

• Text-to-image — FLUX.2 [klein] 9B at 1026 on Artificial Analysis, plus BFL's +91, lands near 1117

• Editing — FLUX.2 [klein] 9B at 1015 on the editing board, plus BFL's +70, lands near 1085

Read that as arithmetic, not as a prediction. Elo deltas are not portable between harnesses. A 91-point gap measured on BFL's prompt set does not transfer point-for-point onto a board built from crowd votes on different images, with different voters and a different sampling window. The exercise is useful only for scale: it suggests a model that would land in the low 1100s in text-to-image, which is roughly where FLUX 3 Image already sits and comfortably behind Nano Banana 2 at 1139. It does not suggest a model that would top the open-weight field by a wide margin, and it certainly does not suggest the +91 translates into first place overall.

If you want the honest one-line version: a low-1100s entry would very likely be the best open-weight model on that board, since the current leader there is 1093. So the direction of BFL's claim is plausible. The magnitude is unverified, and the margin over the nearest open-weight rival may be much smaller than the launch framing implies.

What we saw when we ran it

Our own experience with the model points somewhere more specific than "best open-weight image model." We have run FLUX 3 Image Klein through the vendor's API, without a fixed prompt set or a scoring rubric, so treat this as an observation rather than a result. The capability that stood out was precise, localised editing: changing a single object in a scene, moving an element into a bounded region, or applying an instruction that should leave the rest of the frame untouched. Cases where almost everything is allowed to change — a full restyle, a wholesale lighting change — were much harder to separate from the previous generation.

That is the opposite emphasis from most of the public argument. The launch framing is about whether Klein can edit at all, and whether it beats models that cost five times as much. The more useful question for anyone doing production work is whether it can edit narrowly. Precise edits are where a bounding box and a preserved background actually matter, where a small model either holds the rest of the scene or quietly repaints it, and where the cost of a retry is not just money but a review cycle.

Editing is the claim that deserves the hardest test

Of the three first-place claims, editing is the one to be most sceptical of, for a structural reason. Text-to-image quality is broadly legible — a crowd can look at two images and pick one. Editing quality is conditional on instruction following, and instruction following is exactly the thing a vendor's own prompt set is best positioned to flatter. If your evaluation prompts were written by the people who tuned the model, you are measuring the intersection of two things you control.

There is also a pipeline detail that affects every self-run comparison. Klein ships with a separate prompt upsampler — a 9B fine-tune of Qwen3.5-9B called flux-3-image-klein-pup — which is on by default and which BFL says the model is not meant to run without. That means a naive benchmark that sends prompts as typed is measuring a different system than the one the vendor evaluated. Any independent test worth reading needs to state whether the upsampler was on, because the same checkpoint can produce noticeably different output depending on the answer. That is a lot of methodological surface for a claim that currently rests on one lab's slide.

The speed numbers belong to NVIDIA

One set of figures in the release has a different provenance from the quality claims, and it is worth crediting properly. The latency numbers — roughly 4 seconds for a 1MP image on an RTX 5090 in INT8 at 16 steps, under 9 seconds on an RTX 4070, 0.55 seconds on a GB300 — were measured by NVIDIA, with the RTX results run through ComfyUI. That is still a vendor-adjacent measurement, since NVIDIA helped build the quantized checkpoints, but it is a hardware measurement rather than a preference judgement, and hardware measurements are far easier to reproduce. If you have the card, you can check it in an afternoon.

The 8 GB VRAM claim deserves the same treatment. It is true of the INT8 generator build. It is not true of the whole recommended pipeline, because the prompt upsampler is a separate 18.8 GB BF16 download that wants its own memory. A user who reads "runs on 8 GB" and then enables the default pipeline will find out which claim was scoped and which was not.

How to check this yourself

The cheapest honest test is not a leaderboard. It is twenty of your own prompts, run twice — once with the upsampler and once without — on the tasks you actually care about, scored by whether you would ship the output. Klein makes this unusually practical because the weights are downloadable and the licence permits commercial use below the $5M revenue line, so the test does not require an account with anyone.

For everything else, a routing layer is the cheapest way to run the same prompt across several vendors' endpoints under one key, which is what OrcaRouter is for: one OpenAI-compatible API over 200+ models, 0% markup with the provider's list price passed straight through, and automatic failover when an endpoint degrades. It is worth being explicit that this is not a route to Klein. Our catalogue is 210 models and contains no FLUX entry at all, so for this model the honest options today are BFL's own API or a self-hosted checkpoint. We would rather say that than imply otherwise.

What would settle it

Three things, in ascending order of usefulness. First, Klein appearing as its own entry on an independent board — not merged with FLUX 3 Image, which is a different model at a different price. Second, that entry surviving a few thousand votes rather than a few hundred, because early Elo on a new model is noisy and tends to drift down as the novelty vote fades. Third, someone publishing an editing comparison that states its upsampler setting, because without that the numbers are not commensurable.

Until then the accurate sentence is narrower than the headline and more useful than either extreme. FLUX 3 Image Klein is a 5.6B open-weight image model at $0.009 per 1K image that runs on consumer hardware, whose vendor reports it as the best open-weight model in text-to-image, multi-reference generation and editing. The first half of that sentence is verified. The second half is a claim with a named author, and the author is the lab that built it.