Wan 3.0 is now on Magnific. Thirty seconds of native footage with sound, in a single generation. Up to 10 images, 5 clips and 5 audio tracks as references in one call. Reference, editing and motion control folded into one model.
Most video models hand you a few seconds and let you fight for continuity across the cuts. Wan 3.0 takes a different position: give the model the whole shot, not a fragment of it. Thirty seconds, one generation, no stitching.
That shift moves the hard work somewhere new. When a model can hold a continuous take, the limit stops being the model and starts being the direction. So this post covers both: what Wan 3.0 ships with, and the pre-production that makes a 30-second take worth generating.
The film, start to finish on Magnific. Every shot generated with Wan 3.0.
What is Wan 3.0
Wan 3.0 is the latest generation of Alibaba’s Wan video family, developed by Tongyi Lab and released in public beta on 6 August 2026. Alibaba describes it as an all-in-one reference video generation model.
Three things define it.
Native 30-second generation, with sound. One call produces up to 30 seconds at up to 1080p, audio generated in the same pass, in sync from frame one. Not clips joined afterwards. That is what makes long camera moves and one-take shot language possible, and it doubles the 15-second ceiling of Wan 2.7.
Universal referencing. Up to 10 images, 5 video clips and 5 audio tracks in a single call, combined however the scene needs. Pixel-level consistency holds it together: the character keeps the face, the prop keeps its place, shot after shot.
One model instead of four. Wan 2.7 split text-to-video, image-to-video, reference generation and editing across separate models. Wan 3.0 unifies reference, editing, replication and motion driving, which removes the model-switching step from multi-stage pipelines.
It also proposes a clip length from your prompt instead of making you guess, and extends an existing generation rather than restarting it.
Wan 3.0 vs Wan 2.7: what changed
- Length. 15 seconds becomes 30 seconds, single pass, no stitching.
- Sound. A separate step becomes audio rendered in the same pass, on by default. Switch it off when silence is the point.
- References. Text, image and video becomes up to 10 images, 5 clips and 5 audio tracks in one call.
- Workflow. Four separate models become one, with reference, editing and motion driving unified.
- Duration. You picked blind. Now the model reads your material and proposes the length.
- Aspect ratio. 16:9, 4:3, 1:1, 3:4 and 9:16, plus adaptive, where the model picks the framing from your material.
- Extending. Regenerating becomes video extension from an existing take.
One ceiling worth knowing: Wan 3.0 renders natively up to 1080p. When the take is right, the Magnific video upscaler masters it in 4K without leaving the suite. Nothing gets re-shot.
How to use Wan 3.0 on Magnific
Wan 3.0 is live in the Magnific Creative Suite alongside every other frontier video model, so you can run the same shot through Wan 3.0, Seedance, Kling and Veo and pick the take that works. Billing is by output duration.
Credits per second of output:
- 480p. 100 credits, or 140 on PRIME.
- 720p. 200 credits, or 280 on PRIME.
- 1080p. 400 credits, or 560 on PRIME.
PRIME runs the same specs at the same three resolutions. What changes is speed and latency, so it is the one to reach for when you are iterating fast.
Four steps to your first generation:
- Open the video generator and select Wan 3.0.
- Load your references: start frame, character sheets, style plates, clips or audio.
- Write the shot, not the scene. Camera, subject, movement, duration.
- Generate, then extend rather than restart when a take is close.
That is the mechanical part. The part that decides whether the output is usable happens before any of it.
Direct it like a film, not a prompt
A 30-second continuous take is a directing problem. We learned that making the short film that launches Wan 3.0, produced end to end in Magnific Creative Suite, and the pre-production discipline is the reason it holds together.
Lock the emotional arc first
Decide the arc before anything else. That is what carries the video. In our film the arc is a mother, Susana, stretched thin but present in the moments that count, and her son David, who resents her at times and repays her effort fifteen years later at the premiere of his own film.
The same character at both ends of the arc, fifteen years apart.
Longer shots expose weak intent faster than short ones. Thirty seconds of beautiful nothing is still nothing.
Map the timelines that repeat
If a location appears across scenes set years apart, it has to read as the same place across time. The kitchen in Susana’s flat repeats through the film, so it was locked once and reused everywhere.
The kitchen, generated empty. Every scene set in this room is built on this plate.
That plate is a location, not a shot. No people in frame, every object placed on purpose, so it can be fed back as a reference for every scene set in that room. The prompt is long because the decisions were already made:
LOCATION PLATE PROMPT
35mm film look, visible grain, hazy atmosphere, warm tungsten light. Wide interior location plate of a family kitchen in the morning, no people in the frame, keeping the exact look and atmosphere of the reference image: the same warm ochre walls, the same olive-green painted cupboards with brass knobs, the same mustard square tile splashback, the same freestanding cream enamel cooker, the same bare bulb hanging on a twisted flex, the same small window over the sink with short net curtains and warm morning light coming through, the same open shelves with old tins and jars, the same tea towels on a hook, the same wall calendar, the same lived-in level of clutter, the same wooden worktops. CHANGES from the reference, all of them required: remove the breakfast bar and both wooden stools completely, so that side of the room is open floor. Move the table into the middle of the room, standing free of the walls: a larger plain wooden table for four with a patterned brick-red oilcloth and four wooden chairs around it. At the near end, a child’s place where someone has been making a sandwich: a plate, a sliced loaf, an open jar, a knife. At the far end, directly opposite across the length of the table, a clearly empty chair with a stoneware cup and saucer of coffee in front of it, steam still rising, and lying flat beside the saucer, small and unobtrusive, half a sheet of yellowed paper folded once with a tiny hand-painted boat on it. Remove the pirate hat, the pins, the thimble and any sewing from the room entirely. Remove the printed poster with a ship from the wall. Replace the telephone: a wall telephone in cream plastic with square push buttons and a chocolate-brown handset in its cradle, with a much longer coiled cord hanging in a generous helical spiral that sags well down the wall, long enough to reach across the room. Replace the pot on the cooker: an ordinary cooking pot in dark matte speckled enamel, not polished steel and not shiny aluminium, resting solidly on the cast iron grate with the blue flame visible around its base, filled to the very brim with water trembling, a plume of steam rising, a wooden spoon across the rim. The window frame and the cup are yellowed cream, never bright white. Everything clean, in daily use, cared for: no rubbish, no rust, no cracks, no damp. Details: warm directional morning light from the window mixed with the tungsten of the hanging bulb so the air has body and the haze is visible; dense earthy palette of ochre, olive, mustard, brick red, chocolate brown and yellowed cream, absolutely no white anywhere; matte painted wood, enamel, tile and cloth rather than clean glass or polished steel; moderate depth of field so the cooker, the table and the empty chair all read clearly; open soft shadows, moderate contrast; photochemical grading with fine film grain; no CG render, no miniature diorama look, no plastic sheen, no clean digital rendering, no legible text, no people.
The reverse matters too. The cinema hosting the premiere fifteen years later is a different, real venue, not the school theatre from earlier. A narrative contrast, not a match cut. Decide which spaces echo and which ones deliberately do not.
Lock the look and the lens before generating
Two definitions, written down, applied to every shot: visual style and tone (photographic look, palette, texture, period) and camera language (lenses, distance, movement, depth of field). Both are cheap to define and expensive to retrofit.
Wan 3.0 responds to camera direction across a continuous take, which only helps if you have decided what the camera is doing. If you want to be exact about it, stage the move in 3D Scenes first and hand it to Wan 3.0 as one of its video references. You direct the camera before you spend a single generation.
Lock the cast and the locations
Consistency is a pre-production problem, not a post problem. The film locked eight character references for three characters across every age the story needed: David at 8, 15 and 30, Lucia as a baby and at 7, Susana at 35, 42 and 57. Twelve locations were approved the same way before production started.
Two of the eight locked character references. Approved before a single shot was generated.
With universal referencing, those approved plates become the source of truth you feed the model, shot after shot.
Storyboard, then decide what each shot needs
Thirteen scenes mapped in a shot list. The storyboard tells you which shots need only a start frame and which need a full multishot sequence. With a 30-second ceiling, that call also decides how much of a scene lives inside one generation, and where you chain generations together with the Video Combiner to go long-form.
Build the project like a production pipeline
Create a shared project in Magnific Creative Suite and organize the workflow following the canonical Creative Lab structure. Ours runs through Magnific’s MCP integration in Claude, so folders get created and populated as the film moves.
- Analysis
- Preproduction
- Final Characters: David (8), David (15), David (30), Lucia (baby), Lucia (7), Susana (35), Susana (42), Susana (57)
- Final Locations
- Final Props
- Production
- Final Stills
- Final Clips
- Audio
TIP
Treat your AI workflow like a real film production pipeline. Organization becomes critical as the project grows. Mirror your story structure inside the canonical folders: tag or split key assets by age within the corresponding Preproduction subfolders, so the timelines never get crossed.
Wan 3.0 FAQ
How long can a Wan 3.0 video be?
Up to 30 seconds in a single generation, at 480p, 720p or 1080p. Video extension continues an existing take beyond that.
How many references can I use?
Up to 10 images, 5 video clips and 5 audio tracks in one call, with pixel-level consistency across characters, props, sound and spatial layout.
Does Wan 3.0 come with sound?
Yes. Audio renders in the same pass as the picture and is on by default. Switch it off when silence is the point.
Is Wan 3.0 open source?
No. The weights are closed. Alibaba’s openly released Wan weights stop at Wan 2.2. Any site offering a Wan 3.0 download is not offering Wan 3.0.
Does Wan 3.0 output 4K?
It generates natively up to 1080p. The 4K claims circulating in search results are not from Alibaba. Inside Magnific you finish in 4K with the video upscaler, in the same place you built the references.
What is Wan 3.0 PRIME?
The same model and the same specs at 480p, 720p and 1080p, tuned for speed and lower latency at a different price.
Wan 3.0 or Wan 2.7?
Wan 3.0 for length, continuity and one-take camera work. Wan 2.7 stays available for fast, reliable 1080p generation.
Start directing
The model can hold the shot now. Whether the shot is worth holding is still your call.
Wan 3.0 is available now on Magnific. START MAKING.
