Wan 2.2 A14B

Overview

Wan 2.2 A14B is Alibaba's open-weight mixture-of-experts video family (Apache 2.0). MachGen presents its text-to-video and image-to-video checkpoints as one model: a written prompt creates a new scene, while an attached first frame routes the request to image-to-video.

Both modes create silent five-second 16:9 clips at 480p or 720p with seed control. MachGen serves the checkpoints on managed inference; latency and throughput depend on the selected settings and service load.

Why choose Wan 2.2 A14B

Key strengths

• Open-weight video generation with a mixture-of-experts design focused on cinematic aesthetics and complex motion.
• A unified catalog entry for text-to-video and image-to-video, with the input determining the generation route.
• Seed control makes prompt and composition experiments easier to compare.

Good fit for

• Cinematic concept shots and visual development.
• Animating a still while preserving its main composition.
• Repeatable prompt studies and automated content pipelines.

Inputs and output settings

This table reflects the model routes and controls currently exposed by MachGen.

Using Wan 2.2 A14B

1. Choose a supported generation mode in Playground.
2. Add the required prompt and source or reference assets.
3. Select output settings from the controls shown for that mode.
4. Generate, then inspect the returned asset before reusing the settings through the API.

Practical guidance

• Describe the subject, action or composition, environment, and visual direction in a clear order.
• For image animation, describe the intended motion and camera behavior instead of repeating what is already visible.

Architecture

• Two-expert MoE diffusion. A high-noise expert sketches layout early; a low-noise expert refines detail. About 27B parameters total with ~14B active per step, so cost stays near a dense 14B model.
• SNR-based expert switching. The handoff follows signal-to-noise ratio during denoising.
• Stronger training data than Wan 2.1, with lower validation loss than the non-MoE baseline.

Formats and license

• Text-to-video and image-to-video both output 480p or 720p, 16:9, five-second silent clips at 16 fps.
• Image-to-video uses the supplied still as the first frame.
• Weights are published on Hugging Face under Apache 2.0. Review the model card and applicable terms for your use case.

How to get better Wan 2.2 A14B results

• Keep a five-second shot focused on one primary action and one camera move.
• For image-to-video, describe what should move and how the camera should behave instead of restating the image.
• Use lighting, lens, composition, and color language when a specific cinematic treatment matters.
• Lock the seed while changing one prompt variable at a time, then release it when exploring broader variations.

Limits and availability

• MachGen admits only the modes and settings listed above for this catalog model.
• Source assets must finish uploading before a request can be submitted.
• Generated results can vary between requests, including when the same prompt and settings are reused.
• Wan 2.2 A14B runs on MachGen-managed inference infrastructure rather than being forwarded to a provider API.

Learn more

For model background, technical details, and architecture, see the official Wan 2.2 repository.

Try this model in the MachGen Playground.

Official documentation may describe capabilities or parameters that are not currently exposed by MachGen. Use the Inputs and output settings section above as the source of truth for this page.

Wan 2.2 A14B pricing

See Pricing for currently published MachGen rates. When available, the Playground estimate reflects the selected task and output settings.

Wan 2.2 A14B on MachGen

Deployment: MachGen Hosted. Wan 2.2 A14B runs on MachGen-managed inference infrastructure rather than being forwarded to a provider API.

The Playground and API sections on this page list the inputs and output settings exposed for this model.