Open models, priced per output.

Two image models are live today, z-image-turbo and flux-2-klein-4b, both open-source (Apache-2.0) with a fixed price per output. No video model is live yet. The rest are open-source and open-weight models we don't serve yet. Request early access to start with the live ones.

Request early access Read the docs

Live now.

Two models you can call today, each with a fixed price per output. Every picture is a real output, labelled with its prompt, model and seed.

Z-Image Turbo

Livez-image-turboimage.generate

Fast text-to-image. Up to 4 images per request (num_outputs), PNG, JPEG or WebP.

Takes: Prompt
Returns: 1 to 4 images
Tiers: 1k
Licence: Apache-2.0
From: $0.004 per image (0.08 credits)

Z-Image Turbo details

Editorial fashion photograph of a model in an oversized mustard wool coat on a windy rooftop, overcast sky, strong poseZ-Image Turbo · seed 42

Close-up of a flaky golden croissant on a marble cafe table with a cappuccino, morning light, food photographyZ-Image Turbo · seed 42

Black sand beach with basalt sea stacks under a moody overcast sky, long exposure silky waves, cool muted tones, landscape photographyZ-Image Turbo · seed 11

FLUX.2 [klein] 4B

Liveflux-2-klein-4bimage.generate

The smallest FLUX.2 model, a 4-step distilled 4B text-to-image model. Up to 4 images per request (num_outputs), PNG, JPEG or WebP.

Takes: Prompt
Returns: 1 to 4 images
Tiers: 1k
Licence: Apache-2.0
From: $0.005 per image (0.1 credits)

FLUX.2 [klein] 4B details

1970s Kodachrome photograph of a family on a road trip standing by a station wagon at a desert overlook, warm faded colorsFLUX.2 [klein] 4B · seed 42

Macro wildlife photograph of a hummingbird feeding at a coral hibiscus flower, wings frozen mid-beat, creamy green background bokehFLUX.2 [klein] 4B · seed 42

Skincare bottles in amber glass arranged on travertine blocks with dried palm leaves, warm sunlight and hard shadows, editorial product photography — Skincare bottles in amber glass arranged on travertine blocks with dried palm leaves, warm sunlight and hard shadows, editorial product…FLUX.2 [klein] 4B · seed 11

Image models.

2 are live and priced per output. 6 more are coming soon: models Windpaint doesn't serve yet, listed with their licence and specs but no price. 3 more are listed for reference.

Text to image10 models

Z-Image Turbo
Livez-image-turboimage.generate
Fast text-to-image. Up to 4 images per request (num_outputs), PNG, JPEG or WebP.
Takes
Prompt
Returns
1 to 4 images
Tiers
1k
Licence
Apache-2.0
From
$0.004 per image (0.08 credits)
FLUX.2 [klein] 4B
Liveflux-2-klein-4bimage.generate
The smallest FLUX.2 model, a 4-step distilled 4B text-to-image model. Up to 4 images per request (num_outputs), PNG, JPEG or WebP.
Takes
Prompt
Returns
1 to 4 images
Tiers
1k
Licence
Apache-2.0
From
$0.005 per image (0.1 credits)
Z-Image
Coming soon
The undistilled base of the Z-Image family, meant for fine-tuning and LoRA training.
Developer
Alibaba (Tongyi-MAI)
Takes
Prompt, optional negative prompt
Returns
Images
Parameters
6B
Licence
Apache-2.0
Released
January 2026
Qwen-Image-2512
Coming soon
The December 2025 update of Qwen-Image, a 20B text-to-image model built for rendering text in images.
Developer
Alibaba (Qwen team)
Takes
Prompt
Returns
Images
Parameters
20B
Licence
Apache-2.0
Released
December 2025
FLUX.1 [schnell]
Coming soon
The few-step, Apache-2.0 member of the FLUX.1 family, with a large ecosystem of tools and fine-tunes.
Developer
Black Forest Labs
Takes
Prompt
Returns
Images
Parameters
12B
Licence
Apache-2.0
Released
August 2024
HiDream-O1-Image
Coming soon
An 8B model that works on raw pixels and does text-to-image, editing and subject personalization up to 2048 × 2048.
Developer
HiDream.ai
Tasks
Text to image, Image editing
Takes
Prompt, optional reference images
Returns
Images up to 2048 × 2048
Parameters
8B
Licence
MIT
Released
May 2026
Stable Diffusion XL
Coming soon
SDXL 1.0 base, the latent diffusion model with a large library of community LoRAs and ControlNets.
Developer
Stability AI
Takes
Prompt, optional negative prompt
Returns
Images
Parameters
2.6B UNet
Licence
CreativeML Open RAIL++-M
Released
July 2023
FLUX.2 [klein] 9B
Reference
The larger FLUX.2 [klein] model: text-to-image and multi-reference editing in one 9B model, step-distilled to 4 steps, under a non-commercial licence.
Developer
Black Forest Labs
Tasks
Text to image, Image editing
Takes
Prompt, optional reference images
Returns
Images
Parameters
9B
Licence
FLUX Non-Commercial License
Released
January 2026
Qwen-Image-2.1
Reference
A 7B Qwen-Image release that combines text-to-image, editing with up to 10 references and transparent RGBA output, under a research-only licence.
Developer
Alibaba (Qwen team)
Tasks
Text to image, Image editing
Takes
Prompt, optional reference images (up to 10)
Returns
Images at 2K, optionally transparent
Parameters
7B
Licence
Qwen Research License
Released
September 2026
MiniMax H3-Image
Reference
A 33.1B image model derived from MiniMax H3, doing text to image, up to 16 references and editing at about 16 MP in 6 steps.
Developer
MiniMax
Tasks
Text to image, Image editing
Takes
Prompt, optional reference images (up to 16)
Returns
Images at about 16 MP
Parameters
33.1B
Licence
MiniMax-H3-Image Community License Agreement
Released
November 2026

Image editing1 model

Qwen-Image-Edit-2511
Coming soon
Instruction-based image editing on the 20B Qwen-Image model, with one or more input images.
Developer
Alibaba (Qwen team)
Takes
One or more images and an instruction
Returns
Edited image
Parameters
20B
Licence
Apache-2.0
Released
December 2025

Video models.

0 are live and priced per output. 5 more are coming soon: models Windpaint doesn't serve yet, listed with their licence and specs but no price. 3 more are listed for reference.

Text to video7 models

Wan 2.2 text-to-video
Coming soon
Wan2.2-T2V-A14B makes 5 second clips at 480p or 720p from a prompt alone.
Developer
Alibaba (Wan-AI)
Takes
Prompt
Returns
5 s clip, 480p or 720p
Parameters
27B MoE, 14B active
Licence
Apache-2.0
Released
July 2025
Wan 2.2 TI2V 5B
Coming soon
The small Wan 2.2 model: text-to-video and image-to-video at 720p, 24 fps, in one 5B model.
Developer
Alibaba (Wan-AI)
Tasks
Text to video, Image to video
Takes
Prompt, optional start frame
Returns
Clip at 720p, 24 fps
Parameters
5B
Licence
Apache-2.0
Released
July 2025
LongCat-Video
Coming soon
A 13.6B model for text-to-video, image-to-video and video continuation, built for long clips.
Developer
Meituan (LongCat team)
Tasks
Text to video, Image to video
Takes
Prompt, optional start frame or video
Returns
Clip at 720p, 30 fps
Parameters
13.6B
Licence
MIT
Released
October 2025
HunyuanVideo 1.5
Coming soon
Tencent's 8.3B video model for text-to-video and image-to-video, sized for a single GPU.
Developer
Tencent (Hunyuan team)
Tasks
Text to video, Image to video
Takes
Prompt, optional start frame
Returns
Clip at 480p or 720p, 121 frames by default
Parameters
8.3B
Licence
Tencent Hunyuan Community License
Released
November 2025
LTX-2.5
Reference
Lightricks' 22B audio-video model, with native multishot generation, a Gemma 4 text encoder and an optional duration predictor.
Developer
Lightricks
Tasks
Text to video, Image to video
Takes
Prompt, optional images or video
Returns
Video with synchronised audio
Parameters
22B
Licence
LTX-2.x Community License
Released
August 2026
LTX-2.3
Reference
Lightricks' 22B audio-video model from March 2026, with dev and 8-step distilled checkpoints and a LoRA trainer.
Developer
Lightricks
Tasks
Text to video, Image to video
Takes
Prompt, optional images, video or audio
Returns
Video with synchronised audio
Parameters
22B
Licence
LTX-2 Community License
Released
March 2026
MiniMax H3
Reference
MiniMax's 33B omni video model, generating video with native stereo audio from text, first and last frames or mixed references, at 768p when self-hosted.
Developer
MiniMax
Tasks
Text to video, Image to video
Takes
Prompt, optional first and last frames, or reference images, video and audio
Returns
4 to 15 s clip with stereo audio, 768p
Parameters
33B
Licence
MiniMax H3 Community License
Released
August 2026

Image to video1 model

Wan 2.2 image-to-video
Coming soon
Image-to-video. It animates a start frame you supply.
Developer
Alibaba (Wan-AI)
Takes
Prompt and start frame
Returns
Video clips
Parameters
27B MoE, 14B active
Licence
Apache-2.0
Released
July 2025

Every tier, every price.

Prices are in credits per output. One credit costs $0.05. Aspect ratio and output format don't change the price, and you're charged only for outputs delivered.

Current list prices. The price you pay is the estimate returned at submit, fixed for that job. 1 credit = $0.05.
Model	Capability	Tier	Credits	USD
`z-image-turbo`	`image.generate`	`1k` 1024 px long edge	0.08 per image	$0.004
`flux-2-klein-4b`	`image.generate`	`1k` 1024 px long edge	0.1 per image	$0.005

Ask for the job, not the model.

You submit to a capability such as image.generate. A model implements it, and you can name one or let Windpaint pick.

Submit to a capability
Send POST /v1/generation/capabilities/image.generate with a prompt. Add model to choose one, or leave it out and the capability's first available model runs. The job status reports which model ran.
Pick a tier and a shape
A resolution tier sets one edge and the aspect ratio sets the other. Image tiers set the long edge, video tiers the short edge. All 10 ratios work on every live model.
The price is locked at submit
The estimate is held from your balance and captured when the job completes. Failed, refused and canceled jobs release it. A list price change never touches a job already sent.
Slugs are stable, versions aren't pinned
If Windpaint updates what's behind z-image-turbo, requests that name it get the update. Pair a fixed seed with a versioned preset when you need repeatable output.

Read the catalog.

The catalog endpoint is the source of truth for what's live and what it costs. Read prices and tiers from it rather than hard-coding them.

Catalogbash

# Capabilities, their input slots, and each model's tiers and pricescurl https://api.windpaint.ai/v1/generation/capabilities \  -H "Authorization: Bearer $WINDPAINT_API_KEY"# The same data organized by modelcurl https://api.windpaint.ai/v1/generation/models \  -H "Authorization: Bearer $WINDPAINT_API_KEY"

Each model entry carries available, prices (one row per tier, credits as decimal strings), resolutions, durations, max_outputs, output_formats and licence. POST /v1/generation/estimate validates a request like submit does and returns its price and pixel size without creating a job.

On the schema, not live.

These capabilities are part of the API's schema and appear in the catalog with an empty model list. Submitting to one returns 422. None has a date.

Capabilities without a model. They are listed so you can see the shape of the API, not because they run.
Capability	Inputs	Output	Status
`image.edit`	1 to 3 images, optional mask	image	No model, returns 422
`mask.segment`	1 image	mask	No model, returns 422
`media.upscale`	1 image or video	same as input	No model, returns 422
`audio.speech`	text, optional voice sample	audio	No model, returns 422
`video.lipsync`	1 image or video, 1 audio	video	No model, returns 422
`text.llm`	0 to 8 images	text	No model, returns 422

Questions.

Which model runs if I don't name one?

The capability's first available model in catalog order, so image.generate runs z-image-turbo. Send model: flux-2-klein-4b to use the other image model. If more models ship, the default can change, so name the model when you need a specific one. The job status always reports which model ran.

Can I pin a model version?

No. Model ids are stable slugs, and if Windpaint updates what's behind a slug, requests that name it get the update. Prices are fixed per job when you submit, so a price change never affects a job you've already sent.

Are the models open source?

The live ones are. Both are licensed Apache-2.0, an OSI-approved open-source licence. Some coming-soon models are open-weight instead: the weights are public, but the licence restricts use. The catalog endpoint returns each model's licence, so you can check it from code.

Where do the models run?

On Windpaint's own GPUs. Your prompts and inputs are not sent to a third-party model API to generate these outputs, and Windpaint does not train on your prompts, inputs or outputs.

What does coming soon mean?

A coming-soon model is one Windpaint doesn't serve yet. Its page has the licence and specs from the model's own card, but no model id or price, and you can't call it. None has a date. Request early access to get in while we add them.

What about image editing, upscaling or speech?

Those capabilities are defined in the API schema and appear in the catalog with no models. Submitting to one returns 422. Qwen-Image-Edit-2511 is listed as coming soon for image editing. There's no date attached to any of them.

How do I read prices from code?

Call GET /v1/generation/capabilities. Each model carries a prices array with one row per tier, and credit amounts are decimal strings. POST /v1/generation/estimate prices a specific request and returns the output size without running it.

Open models, priced per output.

Text to image10 models

Image editing1 model

Text to video7 models

Image to video1 model

Submit to a capability

Pick a tier and a shape

The price is locked at submit

Slugs are stable, versions aren't pinned

Start yours.