The mysterious "Happy Horse" suddenly tops the charts, crushing Seedance 2.0. Has video AI changed again?
```
Late Tuesday night, the AI community exploded.
On the globally renowned AI evaluation platform Artificial Analysis' Video Arena leaderboard, a mysterious video generation model codenamed "HappyHorse-1.0" quietly landed in the top spot—no launch event, no technical blog, no company endorsement, soaring to the top with an overwhelming lead.
As of press time, in the text-to-video category, its Elo rating surged to 1357 points, leading Seedance 2.0 (which had topped the list for only five days) by 84 points, and surpassing third and fourth place SkyReels V4 and Kling 3.0 1080p Pro by more than 100 points. HappyHorse-1.0 alone has widened the industry gap.
In the image-to-video category, it achieved a terrifyingly high score of 1402, setting a new record for the leaderboard.
The only area where it fell slightly short was in the overall "video + audio" ranking, where HappyHorse placed second, just below Seedance 2.0.
This leaderboard is not so easy to game
Many people's first reaction: Could this be score manipulation?
This doubt is not unfounded. But Artificial Analysis' ranking mechanism makes it much harder to manipulate than ordinary scoreboards—all rankings come from real global users' "blind two-choice" test votes, where users, without knowing the source, choose between two generated results, and Elo ratings are aggregated from these choices.
The model team cannot cheat by inflating results, so what is reflected is the most authentic perception preferences of everyday viewers.
Of course, some point out that in Artificial Analysis' blind samples, face generation and talking head content account for over 60%, and HappyHorse naturally excels in this area, which might cause some discrepancy between test scores and overall capability.
As a result, discussants on X have split into two camps: skeptics believe there are still obvious gaps between HappyHorse and Seedance 2.0 in character detail and dynamic coherence; supporters see great potential, especially hoping it can solve the industry's pain point of consistent image quality across multi-shot sequences.
Secondly, according to reviews online, ordinary users give this model consistently high ratings.
Whose horse is "HappyHorse"?
This is the question the entire AI community most wants answered.
Speculation on X came quickly. The first thing noticed was the language order on the official website: Mandarin and Cantonese are listed before English. For a product aimed at a global audience, this is quite unusual—it's almost certain the team is from China.
The name itself is also a clue. 2026 is the Year of the Horse in the lunar calendar, and the name "HappyHorse" is an obvious reference; earlier this year "Pony Alpha" played a similar trick. So the suspect list grew quickly: both the founders of Tencent and Alibaba have the surname "Ma" (meaning "horse" in Chinese), so they're natural suspects; some think it might be Xiaomi, since Lei Jun tends to be low-key and likes to surprise everyone; others think the vibe is more like DeepSeek, since DS once quietly launched a visual model and later removed it just as quietly.
X user Passluo commented meaningfully: "Whose HappyHorse is this? Alibaba, Tencent, or Xiaomi?"
Technical "detective work"
Guessing from the name alone isn't enough, so the tech circle kicked into Sherlock Holmes mode.
X user Vigo Zhao compared HappyHorse-1.0's public benchmark data item by item with known models, and found a highly matched candidate: daVinci-MagiHuman—the open source "DaVinci Magic Human" available on GitHub since March this year.
Visual quality, text alignment, physical consistency, and other metrics all match up; the websites are almost identical; both models use a single-stream Transformer architecture, generate audio and video jointly, and have identical supported language lists. Such overlap is hard to attribute to coincidence.
The widely accepted conclusion in tech circles is that HappyHorse is an optimized iterative version of daVinci-MagiHuman by Sand.ai, one of the joint developers, based on the open source model, with the primary goal of verifying upper performance limits in real user preference scenarios, providing groundwork for future commercialization.
daVinci-MagiHuman was officially open sourced on March 23, 2026, as a product of collaboration between two young teams:
One is from the Generative AI Research Lab at Shanghai Chuangzhi Academy, the other is Beijing-based Sand.ai (Sandai Technology). The model uses a 15-billion parameter pure self-attention single-stream Transformer, stuffing tokens for text, video, and audio into a single sequence for joint modeling.
Another lead points to Alibaba Taotian
Meanwhile, another version of speculation is circulating:
The core team behind HappyHorse comes from Alibaba Taotian Group's "Future Life Lab," led by former Kuaishou VP and Kling head Zhang Di.
Public information shows that Zhang Di joined Alibaba at the end of 2025 to lead Taotian Group's "Future Life Lab." This lab is Alibaba E-commerce's core algorithm team, bringing together top technical talent and major compute resources, focusing on large models and multi-modal frontier areas. In just a little over a year, they've published more than 10 high-quality papers at top international conferences.
It's worth mentioning that this rumor started brewing just as Alibaba's Hong Kong stock became more active today—of course, this is just an interesting coincidence. There is currently no concrete evidence directly linking the two, so overinterpretation is not advised.
The real important signal
No matter where HappyHorse ends up belonging, the industry's signal from this incident is already clear enough.
For a long time, there has been a visible gap between open-source video models and closed-source products in terms of quality—when it comes to client delivery, open source models have yet to cross from "usable" to "deliverable." The pricing power of closed-source products like Kling and Seedance is largely based on this gap.
This time, a product based on an open-source model has, for the first time, matched current mainstream closed-source competitors head-to-head in a blind ranking based on real user perceptions.
For closed-source vendors relying on this gap for pricing power, this is at least a signal to take seriously.
According to Artificial Analysis's usual "blind test dominance" routine, when an anonymous model gets enough attention, the official team will usually "claim" it within a week.
Maybe in the next few days, we'll have the answer.
In this Year of the Horse, what really matters may not be which horse runs fastest, but that the track itself is widening.
Risk warning and disclaimerThe market has risks, investments should be made cautiously. This article does not constitute personal investment advice, nor does it consider individual users' specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article are appropriate for their own circumstances. Investing based on this is at your own risk. ```