Alibaba’s AI video model climbs to global top two amid market shakeup

Alibaba Cloud has launched HappyHorse 1.1, a significant upgrade to its AI video generation model, now ranked second globally in independent benchmarks. The model delivers enterprise-ready video synthesis across key content creation scenarios and is accessible via Alibaba Cloud Model Studio with full API support for developers and businesses. A limited-time 40% discount is available for the first two weeks to accelerate adoption.

The timing of this release coincides with rapid shifts in the AI video landscape. OpenAI discontinued its Sora model after deeming it financially unsustainable, while ByteDance paused international expansion of Seedance 2.0 due to copyright disputes with Hollywood studios. These developments have left procurement teams reassessing their generative video solutions, creating an opening for Alibaba to step into the void.

For Alibaba, HappyHorse 1.1 represents more than a technical milestone—it’s a strategic challenge. The model is designed for large-scale integration into enterprise systems, priced competitively and backed by a $52.7 billion global infrastructure. Success in converting technical strength into real-world adoption, especially in Western markets facing U.S.-China technology tensions, could solidify Alibaba’s position in a generative video market projected to exceed tens of billions by 2030.

From unknown entry to benchmark leader in weeks

HappyHorse first appeared in early April as an anonymous submission on the Artificial Analysis Video Arena, a platform where users compare AI-generated videos in blind tests. The model quickly claimed the top spot in both text-to-video and image-to-video categories, later confirmed to be developed by Alibaba’s ATH (Alibaba Token Hub) AI Innovation Unit—a team previously under the Future Life Lab and now part of Taobao and Tmall Group’s restructured divisions.

According to Arena.ai, HappyHorse 1.0 now holds the No. 2 ranking across all three Video Arena leaderboards, scoring 1,444 points in text-to-video and image-to-video categories. It leads Google’s Veo-3.1 (with audio) by 69 points in text-to-video and xAI’s Grok-Imagine-Video by 23 points in image-to-video. In Elo-based ranking systems, these margins reflect consistent user preference, not random variation.

The model’s architecture plays a key role in its performance. Community documentation suggests it’s built on a 15-billion-parameter unified self-attention Transformer that processes text, image, video, and audio tokens in a single sequence. Unlike competitors that rely on separate systems for video and audio, HappyHorse generates all modalities simultaneously, reducing the need for post-processing tools like dubbing or external synchronization software. For enterprises evaluating cost and complexity, this streamlined approach can significantly shorten deployment timelines.

Key upgrades in HappyHorse 1.1 for commercial production

Version 1.1 addresses critical pain points for enterprise video teams, focusing on stability, realism, and workflow efficiency rather than viral social media trends.

• Multi-image reference capability (R2V): Users can upload multiple reference images to maintain consistent character identity across video frames—eliminating the common issue of subject drift seen in earlier AI video models. This feature is essential for brands producing advertising campaigns, product videos, or serialized content where visual continuity is non-negotiable.

• Enhanced motion quality: The upgrade introduces "strengthened motion modeling" to improve fluidity and speed, reducing the unnatural, choppy movements that have plagued AI-generated videos.

• Visual realism improvements: Alibaba targeted persistent artifacts like facial oiliness, over-sharpening, and unnatural textures—flaws that instantly expose machine-generated content. These refinements aim to make outputs indistinguishable from traditional production for critical audiences.

• Advanced audio-visual synchronization: HappyHorse 1.1 introduces "zero-drift lip sync" for dialogue scenes and context-aware speech pacing. The model can now generate up to 15 seconds of 1080p video with fully synchronized audio, a leap from version 1.0’s capabilities.

• Improved instruction following: The update enhances the model’s ability to interpret and execute complex prompts, reducing the need for iterative refinements during production.

Enterprise adoption hinges on trust and integration

Alibaba’s timing may prove advantageous as competitors retreat. OpenAI’s Sora exit and ByteDance’s Seedance delays leave a gap in the market, but winning over enterprise customers requires more than technical benchmarks. Procurement teams demand reliability, scalability, and compliance with regional data governance standards—especially in markets sensitive to geopolitical tensions.

HappyHorse 1.1’s API-first design and enterprise pricing model align with these needs. The model’s unified architecture reduces integration complexity, a major selling point for businesses juggling multiple vendors. However, long-term success will depend on real-world deployment in diverse sectors, from marketing to e-commerce, where consistency and quality are paramount.

As the generative video market evolves, Alibaba’s challenge isn’t just to lead in benchmarks—it’s to prove that its model can deliver production-grade results at scale. If it succeeds, the company could reshape expectations for what AI video can achieve beyond research labs and viral demos.