Skip to main content

Nunchux Optimized models

Nunchux Optimized models run on Nunchux’s proprietary Model Optimizer and Inference Engine. You call them through Nunchux’s own endpoints. Every request is synchronous: you send one POST, and the image or the video clip is in the response. There is no task to poll. All Nunchux Optimized models accept the same request shape. The model field selects the model. On the FLUX, Qwen, HiDream O1, Z-Image Turbo, Krea 2 Turbo and MiniMax H3 Fast models, the tier field selects the tier: Radical Speed or Radical Value. FLUX.1 dev, FLUX.2 dev, Qwen Image and Qwen Image Edit have Radical Speed only. FLUX.2 dev Flash has Radical Value only. Ideogram 4 has three other tiers, Turbo, Balanced and Quality. On Ideogram 4, the extra.rendering_speed field selects the tier. The LTX models have two other tiers, 720p (1280x720) and 1080p (1920x1080). On LTX, the output size (width and height) selects the tier.

Models

Image

Since November 3, 2026, nunchux-qwen-image-2512, nunchux-qwen-image-edit-2511 and nunchux-qwen-image-edit-2509 select the full models. Before that date, these IDs selected the Lightning models. The full models take more steps, so they are slower, and their Radical Speed price is higher. See the pricing page. To use the Lightning models, send the -lightning IDs.

Video

Text-to-image models go to the Text-to-image endpoint. Image-to-image models go to the Image-to-image endpoint. The LTX models go to the Text-to-video and Image-to-video endpoints. MiniMax H3 Fast has its own page, MiniMax H3 Fast, on the text-to-video endpoint.

Choosing a model

FLUX.2 Klein 9B or 4B. The 9B models give higher quality, stronger text rendering, and better results on complex prompts. The 4B models are faster. Use 9B for production output. Use 4B when speed matters more than detail. FLUX or Qwen. FLUX Klein is faster. Qwen Image renders text best, in English and in Chinese, and suits typography and layout work. Qwen Image Edit 2511 Lightning is the recommended editing model when the edit must keep the identity of the subject. HiDream O1. It makes large, photographic images, up to four megapixels, with fine texture in skin, fabric and surfaces. It takes more time per image than FLUX and Qwen. It has a guidance control. HiDream O1 Flash. A distilled HiDream O1 with the same twelve sizes. It is faster, and it takes no guidance and no negative prompt. FLUX.1 dev or FLUX.2 dev. The full-step FLUX models. They take more steps than FLUX.2 Klein, for more detail, and they have a guidance control. FLUX.2 dev is the newer model. FLUX.1 dev also takes a negative prompt, together with true_cfg_scale. FLUX.2 dev Flash. A distilled FLUX.2 dev with the same sizes. It returns an image in about a second. It takes no guidance and no negative prompt, and it has Radical Value only. Qwen Image, Qwen Image 2512, Qwen Image Edit, Qwen Image Edit 2511 and Qwen Image Edit 2509. The full-step Qwen models. They take more steps than the Lightning models, for more detail, and they have a guidance control and a negative prompt. Qwen Image 2512, Qwen Image Edit 2511 and Qwen Image Edit 2509 have both tiers. Qwen Image Edit 2511 and Qwen Image Edit 2509 take up to 3 source images. Qwen Image Edit 2509 and its Lightning model are the earlier release; use them when your work already depends on them. FLUX 3 Image Klein and FLUX 3 Image Klein Edit. The newest FLUX generation. They render in 16 steps, up to 4096x4096 (on Radical Value, at most 4,194,304 pixels, for example 2048x2048 or 4096x1024), and the edit model takes up to 10 source images. They take no guidance and no negative prompt, so write a full description of the scene in the prompt. You can instead send prompt_expansion set to standard, and Nunchux writes the description. FLUX 3 Image Klein Base and FLUX 3 Image Klein Base Edit. The undistilled FLUX 3 Klein models. They take 16 to 50 steps (default 50), a guidance control and a negative prompt, so they give more control and are slower. They render the same sizes, the edit model takes up to 10 source images, and they also take prompt_expansion. Qwen Image 2.1 and Qwen Image 2.1 Edit. They render up to 2752x2752, in sizes that are multiples of 32, and the edit model takes up to 10 source images. They take a negative prompt, which takes effect with true_cfg_scale above 1.0. Ideogram 4. It renders the words in a prompt legibly, for posters, labels and covers. It takes more time per image than FLUX and Qwen. For Chinese text, use Qwen Image. Z-Image Turbo. A step-distilled model that makes photographic images in 8 steps, so it returns quickly. It takes no guidance and no negative prompt, so put everything in the prompt. Krea 2 Turbo. A step-distilled model that makes photographic images in 8 steps, from 1024x1024 up to 2048x2048. Like Z-Image Turbo, it takes no guidance and no negative prompt. FLUX.1 Schnell. The earlier FLUX generation. It stays available for work that already depends on it. LTX 2.5 or 2.3. LTX 2.5 is the newer release, with the same controls as LTX 2.3. Use LTX 2.3 only for continuity with clips you already made. MiniMax H3 Fast. A clip and its soundtrack in one request, up to 15 seconds. Use it when the clip needs sound, or when it must start or end on your frames, or follow reference images, clips and audio. Use LTX for silent clips at 1080p.

Endpoints

The base URL is https://api.nunchux.com. Each page describes one endpoint. The image endpoints also work with the OpenAI SDK. See OpenAI compatibility.

How a request works

  1. Send a POST with a JSON body to the endpoint. Put your API key in the X-API-Key header. See Authentication.
  2. Set model to a model ID from the tables above. Set prompt to the text that describes the output. Image-to-image and image-to-video requests also carry the input image. Then select the tier. Each model uses a different field:
    • FLUX, Qwen, HiDream O1, Z-Image Turbo and Krea 2 Turbo: set tier to radical_speed or radical_value.
    • FLUX.1 dev, FLUX.2 dev, Qwen Image and Qwen Image Edit: set tier to radical_speed.
    • FLUX.2 dev Flash: set tier to radical_value.
    • Ideogram 4: set extra.rendering_speed to turbo, balanced or quality. Do not send steps.
    • LTX: set the size with width and height. The size selects 720p or 1080p.
    • MiniMax H3 Fast: set tier to radical_speed or radical_value, and the size with extra.short_edge and extra.aspect_ratio.
    Do not send tier to Ideogram 4 or LTX.
  3. Read the output from the response. Images return as base64 data or as a URL, selected by response_format. The video endpoint pages describe the video response.
Each endpoint page lists every field and shows a full example.

Billing

Each image request is billed on its tier. Radical Speed gives the lowest latency. Radical Value gives the lowest cost per output. Ideogram 4 bills a fixed price per image, at the rate of its tier: Turbo costs the least and Quality the most. Each LTX request is billed per second of output video, at the rate of its tier: 1080p costs more per second than 720p. Each MiniMax H3 Fast request is billed per second of output video, at the rate of its tier and its short edge, and its reference clips and images can add to the charge. Nunchux charges only for successful requests. When a request fails after the credits are deducted, Nunchux refunds them. See Performance tiers for the tier comparison and Credits & pricing for how credits work.

Errors and limits

A failed request returns an HTTP error status with a JSON body that names the error. See Error codes for the codes and the retry guidance. Every request counts toward the requests-per-minute cap of your plan. See Rate limits.