Alibaba's Wan 2.6 arrived on 16 December 2025 with reference-to-video, which holds a real face and voice steady across scenes and re-dubs a clip into eight-plus languages with phoneme-level lip-sync. It also broke Wan's open-weight tradition. Here is what that means if you market in more than one language.
Wan 2.6 is Alibaba Cloud's video generation model, unveiled 16 December 2025, producing up to 15-second 1080p clips at 24fps with native audio and phoneme-level lip-sync in eight or more languages. Its reference-to-video mode preserves a reference face and voice across generated scenes, enabling localised re-dubs without a reshoot. Unlike Wan 2.1 and 2.2, which were open-weight under Apache 2.0, Wan 2.6 is API-only with no published weights.
For marketers running one concept across several markets, the multilingual re-voice is the draw. The trade is enterprise-style access through Alibaba Cloud and a 1080p ceiling.
| Maker | Alibaba Cloud |
|---|---|
| Released | 16 December 2025 |
| Max clip length | Up to 15 seconds |
| Resolution | 1080p at 24fps |
| Native audio | Yes, multilingual with phoneme-level lip-sync in 8+ languages including English, Mandarin, Japanese, Spanish, French, German, Korean and Russian |
| Modes | Text-to-video, image-to-video, reference-to-video |
| Consistency | Reference-to-video preserves a reference face and, with a voice sample, a consistent voice across scenes |
| Architecture | Mixture-of-Experts, 14B total parameters, about 20% active per inference |
| Weights | Not published. Wan 2.1 and 2.2 were Apache 2.0 open-weight; 2.5 and 2.6 are closed |
| Pricing | Not consistently documented publicly; confirm current rates with Alibaba Cloud Model Studio |
Wan 2.1 and 2.2 were downloadable open-weight releases. Wan 2.5 and 2.6 shipped as closed, commercial API products. 2.6 adds self-insertion of a personal likeness and voice, stronger multi-shot storytelling, and the efficiency-oriented Mixture-of-Experts architecture.
Alibaba Cloud Model Studio (Bailian) API, the official Wan website, and partner platforms such as ImagineArt. No self-hosting.
The multilingual re-voice is the standout: a single ad can be re-voiced and re-lip-synced into eight-plus languages from one reference, which is exactly the job brands running the same TikTok concept across markets pay agencies for.
The catches are structural. It cannot be self-hosted or fine-tuned, unlike its predecessors, access runs through Alibaba Cloud contracts unfamiliar to most Western teams, English documentation is thinner than Google or OpenAI equivalents, and it tops out at 1080p where Kling 3.0 offers native 4K.
Generate anywhere, publish natively to TikTok, Reels and Shorts, attribute signups and sales to every post.
Localise a hero clip in Wan 2.6, then upload each language version to Fastlane and publish natively to the matching TikTok, Instagram Reels and YouTube Shorts accounts, with signups and sales attributed per post and per market. Fastlane also supplies human-warmed accounts in the countries you target.
For everything else on the calendar, Fastlane generates from your website URL: AI UGC videos with 1,000+ characters, slideshows, hook + demo videos and memes. Free plan, no credit card.
Alibaba Cloud's video generation model from December 2025, notable for reference-to-video likeness and voice preservation and phoneme-level lip-sync in eight or more languages.
No. Wan 2.1 and 2.2 were open-weight under Apache 2.0, but Wan 2.5 and 2.6 are API-only with no published weights.
Up to 15 seconds at 1080p and 24fps.
Public per-second pricing is not consistently documented. Confirm current rates directly with Alibaba Cloud Model Studio.
Specs verified against the sources above on September 3, 2026. Model access and pricing change often; check the maker's page before you build on any number here.
Join 50,000+ founders and marketers publishing the same video to TikTok, Instagram and YouTube. Free to start.
