From Happy Horse to Happy Oyster: Inside Alibaba's "Happy Universe" Strategy
Nine days. That's the gap between Alibaba's video model topping every AI video leaderboard and the same team shipping a 3D world model into closed beta. Nine days isn't a release cadence — it's a statement.
On April 7, 2026, a model entered Artificial Analysis's Video Arena anonymously and walked away with the top Elo score in the leaderboard's history: Happy Horse 1.0, a 74-point margin over second-place Seedance 2.0, the largest gap ever recorded on that benchmark. On April 16, the same business group — Alibaba's ATH Business Group — released Happy Oyster, an open-ended 3D world model targeting game content and film production.
Both products live under what Chinese coverage has started calling the "Happy Universe" brand. This piece traces what ATH is actually building, where it fits against Tencent and DeepMind, and what the cadence tells us about what's next.
The Timeline
| Date | Event | Source |
|---|---|---|
| April 7, 2026 | Happy Horse 1.0 appears anonymously on Artificial Analysis Video Arena, climbs to #1 on T2V (Elo 1,347) and I2V (Elo 1,406) | Artificial Analysis |
| April 10, 2026 | Alibaba confirms authorship of Happy Horse 1.0 via CNBC reporting | CNBC |
| April 16, 2026 | ATH Business Group releases Happy Oyster world model; beta waitlist opens at happyoyster.cn | Bloomberg, Cailian Press |
| April 16, 2026 | Alibaba frames Happy Oyster as second release in the "Happy Universe" product line, following Happy Horse | Jiemian News, NetEase |
Two category-defining products in nine days. Happy Horse is a short-clip video model, available on VidCella for 3- to 15-second generations. Happy Oyster is a world model that generates 3-minute interactive 3D scenes. They don't overlap on output — they stack.
Who Is ATH Business Group, and Why Should You Care
ATH Business Group sits inside Alibaba's Taobao and Tmall Group, housed specifically in the Future Life Laboratory (ATH-AI division). According to the Happy Horse 1.0 launch trail, the team lead is Zhang Di, formerly a VP at Kuaishou and one of the principal architects behind Kling AI. His prior work on large-scale video generation explains how a relatively small team has been able to ship two category-defining products at this cadence.
Most observers outside China would have placed Alibaba's serious generative AI work under Qwen or Tongyi. ATH's output reframes that assumption. The team is shipping directly competitive video and world models from Taobao's innovation arm — which tells you something about how much internal budget and infrastructure Alibaba is willing to commit to content generation specifically, versus the core LLM line.
If you want the full technical backstory on Happy Horse 1.0, the review breaking down its 15B single-stream Transformer covers the architecture, the benchmark data, and how it relates to the daVinci-MagiHuman research paper. The open-source access piece is the one to read if you're evaluating whether the promised weights are actually usable yet (spoiler: not quite). Or try Happy Horse 1.0 directly on VidCella — the hosted version is the most reliable access path right now.
The Two-Product Pattern: Clip + Scene
Happy Horse and Happy Oyster are not competing internal projects. They're a stack.
Happy Horse 1.0 takes a prompt, image, reference, or source clip and returns a finished 1080p video with native joint audio and industry-leading lip sync. Its native audio-video architecture treats speech and mouth movement as one token sequence instead of two separate pipelines, which is what produces the category's lowest Word Error Rate of 14.60%.
Happy Oyster takes a different job. Instead of a clip, it generates a 3D environment you can stay inside. Its two modes — Directing for building the scene and Wandering for moving through it — cover a workflow that doesn't exist in a conventional video model. The full Happy Oyster explainer covers the mechanics; the short version is that the model keeps generating as long as you keep interacting, rather than ending after a single prompt.
Stack them together and you get something coherent: a content production pipeline where you prototype the scene in Happy Oyster, render hero shots in Happy Horse or Seedance 2.0, and export the finished piece. That's not two competing products. That's a matched pair.
Where This Sits Against Tencent, DeepMind, and World Labs
Bloomberg's framing of the Happy Oyster release — Alibaba "moving onto Tencent's turf" — tells you most of what you need to know about the domestic competitive picture. Tencent has been pushing 3D content generation for games through its own AI teams; Happy Oyster drops Alibaba into the same category head-first.
Outside China, two efforts are working the same ground. DeepMind's Genie 3 is the most-discussed open-world model internationally, with a focus on playable scenes generated from images. Fei-Fei Li's World Labs, which the South China Morning Post reports is racing Chinese tech giants for the edge in world models, is the other major Western entrant.
What separates Happy Oyster from the Western entrants in its launch messaging: native audio-video joint generation and the explicit Directing/Wandering split. Most world-model announcements so far have focused on the visual environment alone, with audio treated as a secondary system. Alibaba leading with joint sound generation from day one positions Happy Oyster for film and content workflows rather than pure game-engine replacement. That's the kind of decision that reveals go-to-market intent.
What Likely Ships Next
Two product drops in nine days is an aggressive cadence. It also creates some obvious gaps the next release has to fill.
Happy Horse 2.0 or a companion video model. Happy Horse 1.0's hosted range is 3 to 15 seconds at 1080p, with a 5- to 8-second sweet spot. Seedance 2.0 already does 4K and 20-second shots, and the head-to-head breakdown against Happy Horse 1.0 shows where the gaps currently land. The next video release from ATH almost has to push resolution and length without losing the Elo lead.
A Happy Oyster commercial API. Closed beta is a launch marketing choice, not a sustainable access model. Expect an API announcement within a quarter, probably priced per session-minute rather than per-prompt given the streaming generation architecture.
Audio or motion specialist. ATH's edge is native multimodal integration. A dedicated audio-to-video or motion-capture-to-video specialist would slot naturally into the stack without cannibalizing either existing product.
What we'd be surprised to see next from this team: a pure text LLM, or a 3D asset generator aimed at game engines. The content-generation focus so far has been about producing finished or near-finished deliverables, not component assets.
What This Means for Your AI Video Workflow Today
For anyone shipping AI video work right now, the short answer has changed on the video side. Happy Horse 1.0 is now usable through VidCella for hosted text, image, reference, edit, and extend workflows. Happy Oyster is still waitlist-only. The practical production pipeline for April 2026 is Happy Horse for short native-audio clips, with Seedance 2.0 and Wan 2.7 still covering longer shots, 4K-leaning work, and broader API-driven pipelines.
For a longer read on how to choose between the tools that actually ship today, the best AI video models in April 2026 roundup compares Happy Horse 1.0, Seedance 2.0, Veo 3.1, and Kling 3.0 head-to-head. And the Happy Oyster vs Seedance 2.0 breakdown covers exactly where the world-model/video-model boundary lands for production workflows.
Nine days between releases sets a pace. It does not, by itself, open Happy Oyster, but it does make Happy Horse part of what you can ship this week. Track ATH, apply for the beta, and keep your production pipeline split between hosted Happy Horse generation and the models that already have broader APIs.