Motion
The video models, compared
Five engines for turning a script, a frame or a storyboard into motion. They no longer differ on whether they can hold a scene — they differ on how long, how much reference they read, and where the footage has to land.
5 video models from 4 labs · up to 30 seconds a take · native audio on all five
| Model | Max take | References | Audio | Best for |
|---|---|---|---|---|
| Seedance 2.5ByteDance | 30s, extendable | 30 images · 10 video · 10 audio | Native synchronized | One-take storytelling |
| Seedance 2.0ByteDance | 15s | 9 images · 3 video · 3 audio | Native two-channel | Multimodal reference |
| Kling 4.0Kuaishou | 30s (to 2 min) | 10 images · 5 video · 7 subjects | Stereo + lip sync | 4K HDR, long takes |
| Wan 3.0Alibaba | 30s | 20 assets, incl. documents | Dialogue, BGM, SFX | Documents to video |
| MiniMax H3MiniMax | 4–15s | 9 images · 3 video · 3 audio | 32 kHz stereo | Open weights, 2K |
Specs follow each lab's published documentation as of October 2026. Access, resolution and limits vary by region, plan and endpoint.
What changed in 2026
The five models are not iterations of one idea. Together they mark four shifts in what a video model is asked to do.
Thirty seconds, one pass
Seedance 2.5, Wan 3.0 and Kling 4.0 now generate a 30-second take in a single run. Extension carries a sequence from there toward minutes, instead of stitching six-second clips.
References replaced prompts
The input is no longer just a sentence. Seedance 2.5 accepts 50 assets, Wan 3.0 up to 20 plus documents, Kling 4.0 ten images, five videos and seven subjects.
Sound arrives with the picture
Dialogue, music, ambience and effects are generated in the same pass, in stereo. Lip sync is improving with them, rather than being fixed in an editor.
Open weights entered the race
MiniMax H3 ships base checkpoints under a community licence. The strongest open model is now close enough to be a deployment choice, not only a demo.
The five models
Each profile covers what the model is, what it is best at, and a first Skill to run on it.
ByteDance · Seedance 2.5
Seedance 2.5
ByteDance's latest video model: a 30-second audio-video scene from one pass, with up to 50 references.
- Duration
- 30s, extendable
- References
- 50 assets
ByteDance · Seedance 2.0
Seedance 2.0
ByteDance's unified audio-video model: text, image, audio and video references for 15-second multi-shot sequences.
- Duration
- 15s, multi-shot
- References
- 15 assets
Kuaishou · Kling 4.0
Kling 4.0
Kuaishou's production-grade model: 30-second takes, 10-bit HDR, omni-reference and precise keyframes.
- Duration
- 30s
- Output
- 4K HDR
Alibaba · Wan 3.0
Wan 3.0
Alibaba's all-in-one video model: 30 seconds, native sound, and references that include documents and web pages.
- Duration
- 30s at 30fps
- References
- Up to 20 assets
MiniMax · H3
MiniMax H3
MiniMax's open-source omni-modal model: 15-second clips, 2K output and native stereo audio.
- Duration
- 4–15s at 24fps
- Licence
- Open weights
Start with a Video Skill
A Skill is a folder of instructions: the shot list, the pacing, the output format. The model supplies the footage.
Ready to extend your agent?
Connect once, then install any Skill from our studio straight from your agent.