The Reveal
On the morning of April 10, Bloomberg published a brief but significant item: Alibaba's Taotian Group, operating through an internal research division called Future Life Lab under the ATH-AI Innovation umbrella, had quietly built and released HappyHorse-1.0. The Information followed hours later with additional detail on the internal structure of the team and its reporting chain.
Neither story was a surprise to insiders — traces of Alibaba infrastructure had appeared in early community reverse-engineering attempts — but the official confirmation resolves weeks of speculation. The name "HappyHorse" itself had been interpreted as a playful internal codename, consistent with how Taotian's product teams have historically named experiments before productizing them.
What made the release unusual was the complete absence of marketing. Most frontier model launches arrive with blog posts, pre-arranged media embargoes, and demo videos. HappyHorse-1.0 appeared on the Artificial Analysis leaderboard with no announcement at all — and shot to the top of T2V and I2V (no-audio) rankings before anyone knew who made it.
Who Is Zhang Di?
Zhang Di is the technical lead behind HappyHorse-1.0. Before joining Alibaba, he served as VP of Engineering at Kuaishou, where he was the principal architect of Kling AI — the video generation model that briefly held the top spot on Artificial Analysis before being surpassed by newer entrants in late 2025.
Zhang joined Alibaba's Taotian Group in Q4 2025, reportedly recruited directly by Taotian's CTO with a mandate to build a state-of-the-art video generation model from scratch. The move was not publicly disclosed at the time, and his LinkedIn profile still lists Kuaishou as his most recent employer as of this writing.
His background explains several of the model's most distinctive technical characteristics. HappyHorse-1.0's joint audio-video synthesis architecture shares conceptual DNA with the approach Kling AI experimented with internally but never shipped. More strikingly, the model's native Chinese and Cantonese lip-sync capability — which has been widely praised in community testing — appears to draw directly on work Zhang oversaw during his Kuaishou tenure.
The six-language lip-sync support (Mandarin, Cantonese, English, Japanese, Korean, and Spanish) is currently unmatched by any other model at the same quality tier, and it has become one of the features most cited by users who migrate from Kling or SkyReels.
Why Did They Stay Anonymous?
The stealth approach was deliberate. According to sources cited by The Information, the team made a strategic decision to debut on the Artificial Analysis leaderboard without any accompanying announcement. The reasoning was straightforward: a leaderboard position is harder to dismiss than a press release.
The strategy worked. Within 48 hours of the model's appearance, discussion threads on X, Reddit, and Hacker News were already speculating about its origins and testing its outputs. The organic buzz that followed was significantly larger than what a conventional launch would have produced — and it came with a credibility that self-promotion cannot buy.
There is also a competitive dimension. Alibaba's relationship with the AI research community is complicated by its scale and commercial interests. Announcing HappyHorse-1.0 under the Taotian/Alibaba banner from the outset might have attracted skepticism or reflexive comparisons to prior Alibaba AI products. Letting the benchmarks speak first allowed the work to be evaluated on its merits.
What This Means for Open-Sourcing
Since HappyHorse-1.0 first appeared, the team has signaled an intent to open-source the full model — weights, inference code, and training details. The GitHub repository and HuggingFace page have existed since launch but currently show only placeholder content. No timeline has been given.
The Alibaba confirmation changes the calculus somewhat. Alibaba has the infrastructure, legal resources, and precedent (via Qwen and other open releases) to execute a large open-source model release properly. When smaller or more opaque organizations promise open-source releases, the track record is mixed. When Alibaba does, the probability of follow-through is meaningfully higher — though the timeline remains unknown.
There are also commercial considerations. Taotian's core business is e-commerce, and video generation capability has clear applications in product advertising, influencer content, and live commerce. Whether the open-source release will cover the full production model or a smaller research variant is still unclear.
For now, the community best approximation for hands-on experimentation remains daVinci-MagiHuman, the open-source model released by GAIR Lab and Sand.ai in March 2026, which shares architectural similarities with HappyHorse-1.0.