Wan3.0 Video

This LinkModel release covers only wan3.0-video, with text-to-video and first/last-frame generation planned. Multimodal references, editing and extension below describe upstream capabilities and require separate LinkModel production verification before enablement. Prime, file and web-reference inputs are excluded from this release. Identical seeds do not guarantee identical outputs.

Wan 3.0

Wan3.0 is Alibaba's all-in-one video generation and editing model in the Wan family. Its standard wan3.0-video endpoint entered public preview on Alibaba Cloud Model Studio on August 6, 2026. One model covers text-to-video, first-frame and first-last-frame animation, multimodal reference generation, video editing, and video extension, routing tasks from media types and prompt intent.

It generates MP4 videos up to 30 seconds at 30fps in 480P, 720P, or 1080P, with native dialogue, ambient effects, and background music. A request can use up to 20 multimodal references drawn from images, video, audio, documents, and public web pages. The separate wan3.0-video-prime endpoint aligns with the standard model's capabilities while completing generation faster. Each task mode has its own material-combination rules, so input types cannot be mixed arbitrarily.

Key Capabilities

• All-in-One Video Creation: Uses one model for prompt generation, image animation, reference generation, video modification, and extension instead of switching among specialist models.
• 30-Second Multi-Shot Storytelling: Generates up to 30 seconds in one task and supports timestamped sequences of roughly four-to-six-second shots, transitions, and story beats.
• Native Audiovisual Generation: Creates dialogue, background music, and ambient effects alongside visuals, reducing separate dubbing and synchronization work.
• Omni-Reference: Uses up to 20 image, video, audio, file, or web references for characters, objects, environments, motion, sound, and creative information.
• First and Last Frame Control: Animates from a starting image or constrains both the opening and ending composition.
• Prompt-Based Video Editing: Adds, removes, or replaces elements in a reference video and modifies style, lighting, story events, or dialogue.
• Bidirectional Video Extension: Extends footage forward, backward, or in both directions while attempting to preserve format and continuity.

Technical Strengths

Feature	Benefit
Unified Task Routing	Media types and prompt intent select generation or editing behavior, simplifying orchestration for multi-capability products.
30-Second, 30fps Output	A single result can contain a more complete narrative at a frame rate suited to normal playback and editing.
Multiple Resolutions and Smart Duration	Supports 480P, 720P, 1080P and durations from 2 to 30 seconds, with automatic duration selection when requested.
Native Sound Layer	Dialogue, effects, and music are generated with the visuals, improving event timing and audiovisual rhythm.
Heterogeneous Reference Understanding	Media, office documents, and public pages can supply creative context without manually reducing every brief to one prompt.
Standard and Accelerated Endpoints	wan3.0-video and capability-aligned wan3.0-video-prime support validation and faster delivery requirements.

Frequently Asked Questions

How should I write a 30-second multi-shot prompt for Wan3.0?

Begin with the overall theme, visual style, and main characters, then divide the prompt into shot number, timestamp, visual action, and dialogue or sound. Give each shot roughly four to six seconds and specify transitions, camera position, and state continuity instead of placing too many actions in one interval.

How many references can Wan3.0 use, and how should they be assigned?

Wan3.0 accepts up to 20 multimodal reference assets per request, subject to task-specific type and combination rules. Give each asset one role—such as identity, location, motion, voice, or brand information—and refer to it with identifiers such as “Image 1” or “Video 1” to reduce competing signals.

Why can't Wan3.0 use reference audio in first-last-frame mode?

First-frame and first-last-frame modes accept only first_frame and last_frame media types, so reference audio cannot be mixed into those requests. For audio-driven generation, use omni-reference mode with reference_image and reference_audio, then describe how the movement should align with the sound.

How can Wan3.0 edit a video while preserving its original structure?

Supply the source as reference_video, explicitly list what to retain, what to change, and what must not change, and use adaptive format and automatic duration preservation. Start by changing one dimension, such as the background or lighting, before layering additional edits.

How should an existing video workflow connect after Wan3.0 appears on LinkModel?

First verify which task modes, reference combinations, resolutions, durations, audio controls, and editing parameters the LinkModel model page exposes. Submit and retrieve results through LinkModel's documented asynchronous task flow rather than copying Alibaba Cloud workspace fields, headers, or polling behavior.