FLUX.2 [klein] 4B is an open-weight rectified-flow model from Black Forest Labs with a unified architecture. It handles text-to-image generation, single-image editing, and multi-reference generation in a single architecture, and is distilled for 4-step inference.
What it's good at
Klein 4B is designed for workloads where latency matters more than maximum image fidelity:
- Rapid iteration: generate previews quickly while refining prompts, compositions, and styles.
- High-volume workflows: reduce turnaround time across batches and variants.
- Interactive applications: keep user-facing generation workflows responsive.
Speed tiers
Klein 4B runs on two speed tiers, both served by Nunchux's proprietary Model Optimizer and Inference Engine. Select one with the Tier parameter. The tier does not change the model or the settings below. It changes what you optimise for.
- Radical Speed: the fastest generation available, at competitive market rates. Choose it when latency is your binding constraint: interactive previews, or a person waiting on a result.
- Radical Value: the most cost-effective inference, at competitive speed. Choose it when cost per image dominates and you have some latency headroom: large batch jobs and high-volume pipelines.
Advanced settings
- Output size: sets the shape of the output: 1:1 (1024x1024), 16:9 (1280x720), or 9:16 (720x1280). Default: 1:1.
- Seed: fixes the random starting point, so the same prompt, seed, and settings reproduce the same image. Leave it unset when you want a different result each time; if no seed is entered, a random one is chosen for you.
- Inference steps: the number of denoising passes, from 1 to 8. Default: 4. Klein 4B is distilled around a 4-step schedule, so it reaches a good result at the default; adding more steps mostly costs time rather than improving the image.
Practical tips
- Lead with the subject, then add action, style, lighting, and context. Word order matters, because the model weights what comes first. Modelverse does not apply prompt upsampling, so your prompt reaches the model as written.
- Describe the desired result positively. FLUX.2 has no negative-prompt input; use
sharp focus throughoutinstead ofno blur. - Keep the seed fixed while refining the prompt. This isolates prompt changes from seed changes and makes comparisons easier.
