Z-Image Turbo vs FLUX.2 klein (4B and 9B) vs Qwen-Image

At a glance

Text encoder sizes: BFL's launch post gives the 8B Qwen3 embedder for the 9B; the 4B's Qwen3 4B encoder is from musubi-tuner's FLUX.2 docs and matches the repo's config. Z-Image Turbo's and Qwen-Image's encoders are from the configs shipped in their Hugging Face repos.

Architecture and size

Z-Image Turbo uses what Tongyi-MAI calls a Scalable Single-Stream DiT (S3-DiT): text tokens, visual semantic tokens and image VAE tokens are concatenated into one sequence, which the card says is more parameter-efficient than dual-stream designs. It's distilled from the Z-Image base to 8 function evaluations.

FLUX.2 [klein] comes in two sizes, each published as a step-distilled model and an undistilled Base. The 9B pairs a 9B flow model with an 8B Qwen3 text embedder, so in practice it's a much heavier download than the name suggests. A third variant, [klein] 9B KV, adds KV caching; BFL's README describes it as faster than 4B for multi-reference editing.

Qwen-Image is a 20B MMDiT. 2512 is the December 2025 update of the August 2025 text-to-image model; editing lives in the Qwen-Image-Edit line, currently Edit-2511, also 20B.

On raw size, klein 4B is the smallest image model of the three and Qwen-Image the largest by a wide margin, about five times klein 4B and over three times Z-Image Turbo.

Steps, guidance and resolution

These are the settings each card's Diffusers example uses. They're the developers' starting points, not tuned values.

# Z-Image Turbo (ZImagePipeline)pipe(prompt, height=1024, width=1024,     num_inference_steps=9,   # 8 DiT forward passes     guidance_scale=0.0)      # must be 0 for Turbo# FLUX.2 [klein] 4B (Flux2KleinPipeline)pipe(prompt, height=1024, width=1024,     num_inference_steps=4,     guidance_scale=1.0)# Qwen-Image-2512 (DiffusionPipeline)pipe(prompt, negative_prompt=negative_prompt,     width=1664, height=928,  # 16:9 preset     num_inference_steps=50,     true_cfg_scale=4.0)

Step count is the main cost difference when you self-host. klein runs 4 steps, Z-Image Turbo 8 forward passes, and Qwen-Image-2512 50 steps with true CFG. Qwen's README links Qwen-Image-Lightning, a third-party distillation, for fewer steps.

Text rendering

None of these developers publishes a text-rendering score we can compare directly, so here is what each one says about its own model:

• Z-Image Turbo: the card says it "excels at accurately rendering complex Chinese and English text". Its official prompt enhancer puts text to render in double quotes, word for word.
• Qwen-Image: the README describes it as achieving "significant advances in complex text rendering", especially for Chinese. The 2512 card lists improved text accuracy and layout over the August release. Qwen's rewrite guidance quotes every piece of visible text and describes its position, font, colour and size.
• FLUX.2 [klein]: both cards list as a limitation that rendered text "may be inaccurate or subject to distortion". BFL's FLUX.2 prompting guide says to put text in quotation marks and give its placement, style and size.

Z-Image Turbo

FLUX.2 [klein] 4B

If posters, menus or UI mockups with lots of text are your main use, the developers' own positioning points to Qwen-Image first and Z-Image Turbo second. Test with your own copy, especially in languages other than English and Chinese, which none of them makes claims about.

Z-Image Turbo

FLUX.2 [klein] 4B

Z-Image Turbo

FLUX.2 [klein] 4B

Editing: Qwen-Image-Edit vs FLUX.2 klein

Z-Image Turbo doesn't edit. The Z-Image card announces a Z-Image-Edit model, but it's still marked "to be released".

The 2511 card lists what it improved over 2509: less image drift, better character and multi-person consistency, some popular community LoRA effects (lighting, new viewpoints) built in, industrial design generation and geometric reasoning. klein's advantage is that one small checkpoint covers both jobs, so you load one model instead of two.

[klein] 9B edits too, under the non-commercial licence, and the 9B KV variant is the one BFL positions for multi-reference editing.

Licence and VRAM

Z-Image Turbo, klein 4B, klein Base 4B, Qwen-Image-2512 and Qwen-Image-Edit-2511 are all Apache-2.0. klein 9B, 9B KV and Base 9B are under BFL's FLUX Non-Commercial License, which BFL says was renamed from the FLUX [dev] Non-Commercial License without material changes. For a product, that difference usually decides 4B vs 9B before anything else. The klein 4B card also points commercial open-weight users to BFL's usage policy and self-hosted terms, and asks that FLUX marks not imply endorsement.

On memory, klein 4B has two official figures: about 13 GB on the card and launch post, about 8 GB in the GitHub README. Z-Image Turbo fits in 16 GB. klein 9B needs about 29 GB. Qwen publishes no VRAM figure for 2512; its README points to DiffSynth-Studio, a community tool, for low-VRAM offloading. BFL ships FP8 and NVFP4 builds of every [klein] model, and says they use up to 40% and 55% less VRAM.

LoRA training support

• Z-Image Turbo: its family table marks Turbo as not fine-tunable. Train on Z-Image, the undistilled base, which its card positions for LoRA and ControlNet work. musubi-tuner's Z-Image docs recommend the base model now that it's out; before that, they pointed to a de-distilled Turbo from the ai-toolkit author.
• FLUX.2 [klein]: BFL's README says to use the Base models (50 steps) for fine-tuning and LoRA training. Base 4B is Apache-2.0; Base 9B is non-commercial.
• Qwen-Image: no distilled variant, so you train the released checkpoints. The README lists DiffSynth-Studio LoRA and full training as a community option.

Tool support covers all three. ai-toolkit lists Z-Image Turbo, Z-Image, FLUX.2 [klein] Base 4B and 9B, Qwen-Image, Qwen-Image-2512 and the Qwen-Image-Edit releases. musubi-tuner supports Z-Image, FLUX.2 [klein] and Qwen-Image.

Z-Image Turbo vs Z-Image base

Same 6B architecture and licence. Use Turbo for fast, consistent output from a prompt; use the base when you need negative prompts, more variety between seeds, or a model to train on.

FLUX.2 klein 4B vs 9B: which to pick

• Licence: 4B is Apache-2.0; 9B is non-commercial.
• Memory: about 13 GB (RTX 3090/4070 and up) vs about 29 GB (RTX 4090 and up), per the cards.
• Text encoder: Qwen3 4B vs Qwen3 8B.
• Steps: both distilled to 4 steps; both Base versions run 50.
• Capabilities: the same. Both do text to image plus single and multi-reference editing.
• Variants: 9B adds a KV-cached version for multi-reference editing. Both sizes have FP8 and NVFP4 builds.

BFL's README recommends 4B for consumer GPUs. It makes no published quality comparison between the two that we could find, so if you're weighing them for non-commercial work, run both on your prompts.

In short: pick 4B if you'll use the images commercially or have a 16 GB card, since it's the only Apache-2.0 klein and the only one that fits there per BFL. Pick 9B for non-commercial work on a 24 GB-plus card where the larger text encoder is worth the extra memory, or 9B KV if you edit with several reference images. For Z-Image Turbo vs FLUX.2 klein specifically, the split is generation only (Turbo) against generation plus editing in one model (klein).

What about Qwen-Image-2.1 and FLUX.3 Image klein?

Both families had a newer release in autumn 2026.

Qwen-Image-2.1 (September 2026) has a 7B generation component, unifies text to image and editing (up to 10 references, masks, RGBA output) and lists 2048 × 2048 as its 1:1 size. It runs 40 steps; Qwen-Image-2.1-Turbo runs 8 with CFG 1. Against Z-Image Turbo it's closer in size than Qwen-Image-2512 and does more, but it's under the Qwen Research License: non-commercial unless you get a separate licence.

FLUX.3 Image klein (October 2026) is a 5.6B model with up to 10 references and bounding-box guidance. Its card runs the distilled model at 16 steps with no guidance and says to always use its separate 9B prompt upsampler. Peak memory on the card is 31 GB in BF16 at 1024 × 1024 without the upsampler. Its FLUX Kommunity License allows commercial output only for qualifying users under US$5M in annual revenue.

Which one to pick

• You need a commercial licence and fast text to image on a 16 GB card: Z-Image Turbo.
• You need generation and editing from one small model, commercially: FLUX.2 [klein] 4B.
• Your images carry a lot of text, or Chinese text, and GPU memory isn't tight: Qwen-Image-2512.
• You need instruction editing with a dedicated model, commercially: Qwen-Image-Edit-2511.
• You want negative prompts, seed variety or a base to train LoRAs on: Z-Image or FLUX.2 [klein] Base 4B.
• Non-commercial work and a 24 GB+ card: FLUX.2 [klein] 9B is open to you, and so are Qwen-Image-2.1 and FLUX.3 Image klein.

Our hardware requirements and commercial use pages put the published figures and licence conditions for each model in one table.

Trying them without a GPU

Z-Image Turbo and FLUX.2 [klein] 4B are live on Windpaint for text to image, at 1k in ten aspect ratios. Qwen-Image-2512 and Qwen-Image-Edit-2511 are in the catalogue as coming soon.

curl -X POST https://api.windpaint.ai/v1/generation/capabilities/image.generate \  -H "Authorization: Bearer $WINDPAINT_API_KEY" \  -H "Content-Type: application/json" \  -d '{"prompt": "a ceramic teapot on a linen tablecloth, soft window light", "aspect_ratio": "4:5"}'

See the image generation docs for sizes and polling, and pricing for credits.