Alibaba's HappyHorse Just Beat Every Western AI Video Model

The stealth model that quietly topped the global text-to-video leaderboard this week turned out to be Alibaba's. It is 115 Elo points ahead of second place. The top three video models in the world are now all Chinese.

1389
Elo on text-to-video
1416
Elo on image-to-video
+115
Lead over second place
Apr 30
API launch date

Around April 7, 2026, a mystery model named HappyHorse-1.0 appeared on the Artificial Analysis video benchmarking platform, did not identify its creator, and proceeded to climb to the #1 spot on the text-to-video leaderboard. By the time it hit the top, it was leading second-place Dreamina Seedance 2.0 by nearly 115 Elo points — a gap roughly equivalent to the difference between a grandmaster and a solid club player in chess. People started guessing who was behind it. On April 10, Alibaba confirmed: HappyHorse was built by the company's Token Hub innovation unit in Hangzhou.

The second piece of news is actually more important than the leaderboard win. As of April 13, 2026, the top three text-to-video models in the world are all Chinese: HappyHorse (Alibaba), Dreamina Seedance (ByteDance), and Kling (Kuaishou). Sora, Veo, and Runway — the three US names people associate with AI video — are all trailing. This is the first time in the history of the category that no American model holds the #1 slot.

The 5-Second Version

01

The Leaderboard Story

Artificial Analysis runs blind-test leaderboards for text-to-video and image-to-video models. Users submit prompts, the platform shows them two randomly-selected outputs from different models, and users pick the better one. The results are Elo-rated like chess. It is the closest thing to a neutral global leaderboard that video AI has.

Around April 7, a model labeled only "HappyHorse-1.0" showed up. No one knew who built it. Over three days it steadily climbed until it passed every model in both categories. Here is roughly where the top of the leaderboard stands as of April 13:

#
Model
Org
Elo
1
HappyHorse-1.0 (Alibaba)
CN
1389
2
Dreamina Seedance 2.0 (ByteDance)
CN
1274
3
Kling 2.5 (Kuaishou)
CN
1241
4
Veo 3 (Google DeepMind)
US
1218
5
Sora 2.1 (OpenAI)
US
1205

Three Chinese models before the first American one. First time ever.

02

Who Actually Built This

Alibaba Token Hub is the innovation unit within Alibaba's DAMO Academy that has been quietly shipping video generation work for the last 18 months. They released an earlier model called Tongyi Wanxiang in mid-2025 that was respectable but not frontier. HappyHorse-1.0 is their next-generation effort, and based on the leaderboard results, it is a major step forward.

The model appears to be a diffusion-transformer hybrid trained on a mix of licensed video data and Alibaba's own e-commerce/entertainment corpus. It supports all four major video generation modes: text-to-video with and without native audio, and image-to-video with and without native audio. That is unusually comprehensive — most Western video models still handle audio as a separate post-processing step.

Bo's honest read: the gap is probably somewhat overstated by the Elo number. Benchmarks swing hard when a stealth model launches because early voters are enthusiasts who get excited. The gap will narrow as more people vote, but even if it compresses to half — HappyHorse is still clearly at or near the frontier, and Alibaba did not used to be on the video map at all.

03

What This Actually Means

01

China Is Winning at Open-ish Video AI

Three Chinese labs in the top three slots is not a fluke. ByteDance, Kuaishou, and now Alibaba have been systematically shipping video model improvements while Sora hit pause post-controversy and Runway focused on monetization. The Chinese labs have the training data advantage (TikTok, Douyin, Kuaishou, Taobao video) and the research talent, and they have been moving faster.

The gap is structural, not temporary
02

Creators Should Already Be Testing This

If you make video content for a living — marketing, ads, YouTube, TikTok — the best tool on the leaderboard is worth testing the week it launches. The API opens April 30. Don't wait for the "official" English-language tutorial economy to catch up. Sign up, run your actual workflows through it, see if it replaces or supplements your current stack.

Test new models within 2 weeks of launch
03

Western Labs Are Not Out of the Race

Sora 2.0 was a big jump and Veo 3 is genuinely good. OpenAI has been conservative post-2024 around video for safety reasons, which has slowed their leaderboard climb but not their underlying capability. Expect a Sora 3 or Veo 4 release before end of 2026 that is very competitive with HappyHorse. Betting the US labs are permanently behind is premature.

Don't short the US labs yet
04

Chinese API Access Is a Real Question

If HappyHorse's API is only available through Alibaba Cloud's China endpoint, US businesses will face data sovereignty and compliance headaches that make adoption hard. If Alibaba exposes it through Alibaba Cloud International, the friction drops substantially. Watch the April 30 launch details carefully — the deployment architecture will tell you how serious Alibaba is about Western adoption.

The access path matters as much as the model
04

The Bottom Line

The Verdict
HappyHorse is a real frontier video model and Alibaba is now a serious video-AI lab. The leaderboard gap will narrow but the capability jump is genuine. If you produce video content for work, April 30 is a date to mark on your calendar.

The broader story is that video AI is moving the way image AI moved between 2022 and 2024: three or four labs trading the #1 spot every few months, with the top of the leaderboard becoming indistinguishable from professional work. For the next six months, video is the most dynamic AI category in the world, and the winners are going to be the creators who test every new release the week it ships, not the ones waiting for the category to stabilize.

Stop Reading AI News. Start Building With It.

The 2-day in-person Precision AI Academy bootcamp. 5 cities. $1,490. 40 seats max. Thursday-Friday cohorts, June-October 2026.

Reserve Your Seat
PA

Published By

Precision AI Academy

Practitioner-focused AI education · 2-day in-person bootcamp in 5 U.S. cities

Precision AI Academy publishes deep-dives on applied AI engineering for working professionals. Founded by Bo Peng (Kaggle Top 200) who leads the in-person bootcamp in Denver, NYC, Dallas, LA, and Chicago.

Kaggle Top 200 Federal AI Practitioner 5 U.S. Cities Thu-Fri Cohorts