FLUX 3 Image Klein vs Qwen-Image 2.1 Turbo: The Licence Decides This One Before the Benchmarks Do

FLUX 3 Image Klein, from Black Forest Labs on 15 October 2026, and Qwen-Image 2.1 Turbo, from Alibaba a week earlier on 9 October 2026, look like the same product from two companies: small open-weights image generators that handle text-to-image, editing and multi-reference composition in one set of weights, both accepting up to ten reference images, both shipping with accelerated inference paths. Put the spec sheets next to each other and Qwen-Image 2.1 Turbo wins on most rows — more parameters, fewer denoising steps, transparency, masks. Then read the two licences, and the smaller, slower-looking, more expensive-to-reason-about model turns out to be the only one of the two you can build a business on without signing anything.

That inversion is the whole comparison. Everything else — the parameter counts, the prices, the render times — is a detail you weigh once you have cleared it.

What each model actually is

FLUX 3 Image Klein is a 5.6B-parameter image transformer under the FLUX Kommunity License, released as open weights plus a metered endpoint on BFL's own API starting at $0.009 per 1K image. It is the smaller member of the FLUX 3 Image family and inherits that family's surface: native 4K output, bounding boxes written as [top, left, bottom, right] on a 0–1000 grid, and up to ten reference images. It ships in two checkpoints — a step-distilled build at 16 steps for generation, and an undistilled base at 50 steps for fine-tuning and LoRA training — plus FP8, INT8 and NVFP4 quantisations. The 5.6B counts the image transformer alone; the text encoder, Qwen3-VL-4B-Instruct, is a separate download.

Qwen-Image 2.1 Turbo is an accelerated checkpoint of Qwen-Image 2.1, keeping the same 7B visual-generation architecture — 32 single-stream DiT layers — and cutting inference to 8 denoising steps with a saved sampling schedule and prefix KV caching. It is a unified generation-and-editing model with a feature list Klein does not match: native transparency, so it generates and edits RGBA images directly; and local edits through drawn circles, painted annotations or separate masks rather than bounding boxes. Weights are under the Qwen Research License Agreement, and the hosted version is priced at $0.016 per image through Alibaba Cloud Model Studio and QwenCloud.

Both are true open-weights releases in the sense that matters to a downloader: you can get the files, run them locally, and fine-tune them. What you may then do with the output is a different question, and it is the one the two vendors answer differently.

The licence is where they part company

The Qwen Research License Agreement is unambiguous at the top: the grant is "FOR NON-COMMERCIAL PURPOSES ONLY", and section 2(b) says you "shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us". The word the agreement uses for Non-Commercial is "research or evaluation purposes". That means the Qwen-Image 2.1 Turbo weights cannot go into a product without a negotiated licence from Hangzhou Tongyi Laboratory. Buying the model through Alibaba's hosted API is a commercial service and carries its own terms — the restriction is on self-hosting the checkpoint, which is exactly what a team picks an open-weights model to do.

FLUX 3 Image Klein goes the other way. Under the FLUX Kommunity License v1.0 the weights are free for non-commercial use, and there is a commercial-output exception for "Qualifying Users" — individuals and companies under US$5M in gross annualized revenue. Below that line you can ship Klein's output commercially at no cost beyond your own compute. Above it, you need a self-hosted licence from BFL or you use the API.

So the split is not "one is open and one is closed". Both are downloadable. One is downloadable for research, the other is downloadable for a business up to a revenue threshold. For an indie developer or a small studio, that single clause decides the matchup outright.

Two further terms are worth knowing before you commit either way. Klein's licence forbids using the model, its derivatives or its outputs to improve another AI model performing a similar function — synthetic-data generation for training a rival image model is the case that clause covers. And Qwen's agreement requires redistributed copies to carry its attribution notice and any modified files to be marked as changed, which is ordinary open-source bookkeeping but is a real obligation if you plan to publish a fine-tune.

Specs, row by row

• Image generator — FLUX 3 Image Klein 5.6B parameters vs Qwen-Image 2.1 Turbo 7B (32 single-stream DiT layers), each counted the way its publisher counts it, text encoders excluded from both.

• Denoising steps — Klein distilled at 16 steps, base at 50 vs Qwen Turbo at 8, with its sampling schedule saved inside the checkpoint.

• Max output — Klein up to 4096px-class native 4K vs Qwen Turbo's published presets topping out at 2752×1536 (about 4.2 MP), with 2048×2048 the largest square.

• Reference images — ten on both.

• Region control — bounding boxes on a 0–1000 grid for Klein (placement and scale, explicitly not a hard pixel mask) vs drawn circles, painted annotations or a separate mask for Qwen-Image 2.1.

• Transparency — Qwen-Image 2.1 generates and edits native RGBA; FLUX 3 Image Klein makes no transparency claim on either its model card or its product page.

• Prompt handling — Klein requires a prompt upsampler in front of it and its model card says outright not to run it without one; Qwen Turbo takes the prompt as written, with prompt extension available as an API parameter.

• Licence — FLUX Kommunity License, commercial output free under $5M revenue vs Qwen Research License, non-commercial only, separate licence required to use the weights commercially.

• Vendor API price — $0.009 per 1K image, rising to $0.036 at 4K vs a flat $0.016 per image at every preset.

What the two price lists actually do at volume

Klein's rate card scales with resolution; Qwen Turbo's does not. That produces a crossover that is easy to miss when you compare the headline numbers.

• 10,000 images at 1K — Klein $90 vs Qwen-Image 2.1 Turbo $160. Klein is 44% cheaper.

• 10,000 images at 2K — Klein $140 vs Qwen-Image 2.1 Turbo $160. Klein is still cheaper, by 12.5%.

• 10,000 images at 4K — Klein $360. Qwen-Image 2.1 Turbo has no 4K preset to price against; its largest is 2752×1536 at the same $160.

Those are arithmetic on two published rate cards, not quotes. The crossover sits between 2K and 4K, and it exists because the two vendors priced in different units: BFL charges per image by resolution band, Alibaba charges per image flat. If your workload lives at 1K or 2K, Klein is the cheaper call on both APIs, which is the opposite of what "the cheap Qwen model" would lead you to expect. If it lives at true 4K, Qwen-Image 2.1 Turbo cannot serve it at all through the published presets, so the question is not price.

The comparison gets one more wrinkle when you self-host instead of calling an API. On your own hardware there is no per-image rate at all — there is a fixed cost for the GPU and a per-image cost in seconds — and the two models are not close in what they need to run.

Running them yourself

Sum the published file listings and the gap is stark. FLUX 3 Image Klein's three BF16 components — an 11.1 GB transformer, an 8.9 GB text encoder and a 0.34 GB autoencoder — come to roughly 20.3 GB, and the INT8 build runs on an RTX with as little as 8 GB of VRAM, with BFL's documentation putting the floor at an RTX 3090. Qwen-Image 2.1 Turbo's BF16 download is dominated by a 17.53 GB text encoder, with the transformer split across a 9.97 GB and a 4.26 GB shard and a 0.68 GB VAE — about 32.4 GB in total, so its text encoder alone is larger than everything Klein needs put together. Both figures are our sums of the vendors' own file listings, not vendor-published totals.

Speed is where the step counts mislead. Klein's distilled checkpoint runs twice as many steps as Qwen Turbo's eight, but BFL's measured numbers — taken by NVIDIA, in ComfyUI, on the INT8 distilled build — put a 1 MP image at about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070, dropping under 0.6 seconds on a GB300 in NVFP4. Alibaba does not publish a comparable per-GPU figure for Turbo, so the honest position is that Turbo's eight steps are architecturally the shorter path and nobody outside Alibaba has published seconds for it on consumer hardware.

Budget for Klein's upsampler, whichever model you pick. The prompt upsampler is not optional for it — flux-3-image-klein-pup is a separate 9B vision-language model, an 18.8 GB BF16 download that adds around 28 GB of GPU memory on an 80 GB card when it runs alongside the generator. That single dependency erases Klein's VRAM advantage for anyone deploying on one small card, unless the caption is produced elsewhere and passed in.

Neither model has an independent quality score

This is the row both spec sheets leave blank, and it is worth being precise about how blank. As of 16 October 2026, Artificial Analysis runs separate text-to-image and image-editing boards, and neither carries an entry for FLUX 3 Image Klein or for Qwen-Image 2.1 Turbo. What the boards do carry is the previous generation from each family, which is the closest thing to an independent read available:

• Text-to-image — Qwen-Image 2.1 (the base checkpoint, not Turbo) ranks 20th at 1,041 Elo from 5,356 comparisons, released September 2026. FLUX 3 Image, the larger sibling Klein sits under, ranks 13th at 1,114 Elo from 5,769 comparisons.

• Image editing — Qwen-Image 2.1 ranks 19th at 1,081 Elo from 6,138 comparisons. FLUX 3 Image has no entry on the editing board at all; the nearest FLUX model listed is FLUX.2 [Klein] 9B at 1,015 Elo from 6,905 comparisons.

Read those as context, not as a prediction. The base Qwen-Image 2.1 is not Turbo — a distilled checkpoint trades quality for steps, and how much is exactly the question Turbo's absence from the board leaves open. FLUX 3 Image is not Klein either, and Black Forest Labs' own launch post never compares the two, which means there is no published figure anywhere stating how far Klein falls below its bigger sibling. And on the editing board, the model BFL says Klein beats by 70 Elo — FLUX.2 [klein] 9B — is sitting at 1,015, roughly 66 Elo below where Qwen-Image 2.1 lands. BFL's claim is that Klein clears that gap and more; the only evidence for it right now is BFL's own chart.

Which one to pick

Choose FLUX 3 Image Klein if the output has to be commercial and the business is small. It is the only one of the two whose weights come with a usable commercial path attached without a negotiation, it is cheaper per image on both APIs at 1K and 2K, and it reaches 4K. Budget for the upsampler, and read the licence's restrictions on competitor-training and fine-tune redistribution before you build a data pipeline around it.

Choose Qwen-Image 2.1 Turbo if transparency or mask-based local editing is the feature you actually need — Klein offers neither — or if you are running research, evaluation or an internal prototype where the non-commercial clause never binds. Eight steps and a 7B architecture make it an efficient thing to experiment with locally; the 32.4 GB BF16 footprint is the cost of that experiment.

If you need both, the split is legitimate: Qwen-Image 2.1 Turbo for the exploration phase and its transparency path, FLUX 3 Image Klein for the shipping path. That is a licensing decision as much as a technical one.

Where to actually call them

Both models are reachable through the vendor's own API — BFL for Klein, Alibaba Cloud Model Studio and QwenCloud for Turbo — and through several third-party platforms. Neither is in OrcaRouter's catalogue today, and we would rather say so than gesture at a routing story we cannot back. What we can say is where the surrounding workflow fits: OrcaRouter already routes a set of hosted image models alongside 200+ text and multimodal models behind one API key, at 0% markup on the provider's list price with automatic failover between providers. Teams using a hosted image model for the interactive path and keeping an open-weights checkpoint for batch work are running exactly the split a single endpoint is built for, and if either of these two lands in the catalogue, that is where it will appear.

The thing to watch is not the next spec sheet. It is the first independent Elo for either model, on the same board where their predecessors already sit. Until that lands, the two claims in play — BFL's "leading open weights quality with fewer parameters" and Alibaba's "most balanced and cost-effective" — are both advertising, and the only rows in this comparison you can act on with confidence are the licence, the price list and the VRAM floor.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.