The global race for generative video has a fresh, unexpected frontrunner. In a move that has caught the AI community off guard, a model dubbed HappyHorse-1.0 appeared in early April, quietly scaling the leaderboards of Artificial Analysis to claim the top spot across four major categories. The model didn’t arrive with a flashy keynote or a corporate press release; instead, it emerged as a mystery, listed simply as “coming soon” even as it systematically outperformed established giants like ByteDance’s Seedance 2.0 and Kling.
The silence ended on April 10, when Alibaba’s ATH division confirmed that HappyHorse is an internal research project currently in closed beta. The stakes for this release are high: HappyHorse-1.0 is positioned as the first open-source video model capable of natively synchronizing audio and video generation. With a public API scheduled for release on April 30, Alibaba is shifting from a quiet launch to a strategic offensive in the multimodal AI landscape.
This pattern of “silent deployment” followed by a major announcement is becoming a signature move for Chinese AI firms. Similar tactics were recently seen with Xiaomi’s “Hunter Alpha” and Zhipu AI’s “Pony Alpha,” suggesting a broader industry shift toward letting benchmark performance create demand before the official marketing machine kicks in. For Alibaba, the timing is precise, arriving just as the industry questioned the momentum of OpenAI’s Sora, effectively signaling that the torch of video innovation has moved East.
Breaking the Benchmark: Visual Dominance and Audio Parity
The impact of HappyHorse-1.0 is most evident in the Elo ratings provided by Artificial Analysis. As of April 13, the model achieved a score of 1,384 in text-to-video (without audio), leading Seedance 2.0 by a substantial 111 points. Even more striking was its performance in image-to-video, where it hit 1,413 Elo—the highest score ever recorded on the platform.
In the world of Elo ratings, a gap of 60 points typically indicates a clear user preference. A 111-point lead suggests that in blind tests, users overwhelmingly preferred the visual output of HappyHorse over its closest competitors. The model is particularly praised for its rendering of human figures, with testers noting superior skin textures and fluid movement, especially in portrait and talking-head clips, which made up over 60% of the test samples.
But, the dominance is not absolute. When audio is introduced into the equation, the gap narrows significantly. In categories involving synchronized sound, HappyHorse and Seedance 2.0 are essentially tied, with only 1-2 points separating them. This indicates that while HappyHorse has mastered the visual realm, its audio synchronization and sound quality are on par with, rather than superior to, the current industry standard.
The Architecture of a ‘One-Pass’ Model
What separates HappyHorse-1.0 from its peers is a fundamental departure in how it “thinks” about video. Most current AI video tools use a pipeline approach: they generate a silent video first and then layer audio on top using a separate process. HappyHorse utilizes a Unified Transformer architecture consisting of 40 layers of self-attention, allowing text, video, and audio tokens to be processed in a single sequence.
By modeling sound and imagery in the same semantic space from the start, the model achieves a “one-pass” generation. This allows for native lip-syncing across seven languages, including English, Mandarin, Cantonese, Japanese, Korean, German, and French. The technical efficiency is further bolstered by DMD-2 distillation and MagiCompiler optimization. In a controlled environment using a single H100 GPU, the model can generate a 5-second 1080p video in approximately 38 seconds.
| Feature | HappyHorse-1.0 | Seedance 2.0 |
|---|---|---|
| Access Model | Open Source | Closed Commercial |
| Architecture | Unified Transformer | DB-DiT (Bidirectional Diffusion) |
| Generation Mode | One-pass (Simultaneous) | Pipeline-based |
| Max Output | 5–10 seconds (1080p) | Up to 60 seconds (2K) |
| Parameters | 15 billion | Not disclosed |
Despite the technical triumphs, the model is not without flaws. Leaked clips have revealed “rippling” artifacts in certain frames and streaking on fast-moving objects, with some users noting a dip in quality when the video is viewed on larger screens.
A comparison provided by Artificial Analysis showcases the model’s ability to handle complex prompts, such as a Pixar-style animation of a timid traffic cone dreaming of becoming a finish-line marker, complete with a transition from traffic noise to cheering crowds.
The Strategic Pivot: From Lab to Market
The emergence of HappyHorse is a direct result of a massive organizational reshuffle within Alibaba. The project is housed within the AI Innovation Department of the Alibaba Token Hub (ATH), a group established on March 16 under the direct leadership of CEO Wu Yingming. ATH was designed to integrate various arms of the company—including the Tongyi Lab and the Qwen team—to accelerate the transition from basic research to commercial application.
Central to this success is Zhang Di, a heavyweight in the AI space often referred to as the “Father of Kling.” After serving as Vice President at Kuaishou and leading the technical development of the Kling AI model, Zhang joined Alibaba in November 2025 to lead the “Future Life Laboratory.” Within five months of his arrival, his team produced HappyHorse-1.0, effectively beating his former employer’s technology at its own game.
The market has responded with enthusiasm. Following the confirmation of the project, Alibaba’s stock saw a surge of over 7%, closing at 126.6 HKD on April 10. This success creates a “dual-engine” structure for Alibaba: Tongyi Lab focuses on the deep, foundational research, while the ATH Innovation Department builds applications tailored to real-world business challenges.
Access and Implementation
For developers and creators, the most critical detail is the open-source nature of the project. On April 9, the weights were made public via GitHub without commercial restrictions. However, the barrier to entry remains the hardware. With 15 billion parameters, running the model locally on consumer-grade hardware like an RTX 4090 (24GB VRAM) requires 4-bit quantization, which some users report leads to a noticeable drop in quality. For professional-grade output, cloud GPUs with at least 40GB of VRAM are recommended.
Users are cautioned to be wary of phishing attempts; the official team has warned that many “official” websites currently circulating are fraudulent. The legitimate portal is still being finalized, and the most reliable source of truth remains the official GitHub repository and ATH announcements.
The next major milestone for the project is the public release of the API on April 30, which will allow users to integrate HappyHorse’s capabilities without the need for massive local compute resources. Whether this open-source approach will force other commercial giants to open their gates remains to be seen.
Do you think open-source models will eventually replace closed-system AI video generators? Share your thoughts in the comments below.
