Skip to main content
Back

Alibaba's HappyHorse 1.1 Hits #2 in Video AI While Rivals Stumble

Alibaba's HappyHorse 1.1 lands at #2 on the Artificial Analysis video leaderboard, as OpenAI exits and ByteDance stumbles, reshaping who controls AI video generation

  • HappyHorse 1.1 ranks #2 on the Artificial Analysis Text-to-Video leaderboard with an Elo of 1,152, behind only ByteDance's Seedance 2.0 (1,219).
  • Biggest gain: Image-to-Video with Audio jumped from #5 to #2, driven by stronger audio-visual synchronization.
  • Architecture: A 15B-parameter unified self-attention Transformer generates video and audio in a single pass, with 8-step inference via DMD-2 distillation.
  • New in 1.1: Native lip-sync in 7 languages, up to 9 reference images, 720p and 1080p output, improved motion and scene consistency.
  • Pricing: $9.90/min at 1080p, available on Alibaba Cloud Model Studio, Qwen Cloud, and fal.ai with a 40% launch discount.
  • Market context: OpenAI shut down Sora; ByteDance paused Seedance 2.0's international rollout over Hollywood copyright disputes, opening the door for Alibaba.

The AI video generation race just got a clearer picture at the top. Alibaba's HappyHorse 1.1 has debuted at #2 on the Artificial Analysis Text-to-Video leaderboard, sitting just behind ByteDance's Seedance 2.0 with an Elo score of 1,152 versus Seedance's 1,219. It is a strong showing for a model that only entered the market a few months ago, and it arrives at a moment when the competitive landscape around it has shifted dramatically in Alibaba's favor.

A model born in stealth

HappyHorse's origin story is unusual. HappyHorse 1.0 appeared on the Artificial Analysis benchmarking platform around April 7, without identifying its affiliations, and climbed to the top of blind-test rankings for both text-to-video and image-to-video generation. The anonymous debut sparked widespread speculation before the developers revealed on X that HappyHorse was part of Alibaba's ATH AI Innovation Unit, and Alibaba confirmed to CNBC that the post was genuine. Alibaba's Hong Kong-listed shares closed 2.12% higher the day the news broke.

Version 1.1 is a refinement, not a rebuild. Compared with version 1.0, HappyHorse 1.1 delivers upgrades across several key areas, including motion dynamics, subject consistency, prompt adherence, visual quality, and audio generation capabilities. Alibaba's biggest push with this release is on audio-visual synchronization, and the Artificial Analysis data backs that up: the model's largest gains over 1.0 come in the Image-to-Video with Audio category, where it jumped from #5 to #2.

What's under the hood

The architecture powering HappyHorse is worth understanding because it is meaningfully different from most competing models. HappyHorse is a roughly 15B-parameter unified self-attention Transformer. Where most video models use dedicated cross-attention branches to inject text conditioning and separate audio modules entirely, HappyHorse concatenates text, image, video, and audio tokens into a single sequence, and the same attention layers process everything.

The layer layout is a sandwich: the first 4 and last 4 layers handle modality-specific projections, while the middle 32 layers share parameters across all modalities. Audio-video alignment is learned as part of denoising rather than added as a post-processing fix. Inference uses DMD-2 distillation (a technique that compresses many diffusion steps into very few) at eight sampling steps, with a reported wall-clock time of roughly 38 seconds for 1080p on an H100.

The practical result of this design is that you get video and audio in a single generation pass. Key capabilities in 1.1 include:

  • Native audio with lip-sync in seven languages, including English, Mandarin, Cantonese, Japanese, Korean, German, and French
  • Up to nine reference images to anchor characters, environments, or visual style across a multi-scene project
  • 720p and 1080p output across text-to-video, image-to-video, and reference-to-video workflows
  • Stronger motion expressiveness and scene consistency over 1.0

One notable limitation: the lip-sync is driven by HappyHorse's native audio generation, meaning the model creates audio that matches the video it generates. You cannot feed in an existing audio file and have characters sync to your voiceover. This is the most-requested missing feature in early feedback, and given that joint audio-video is already in the architecture, audio conditioning from user input seems like a near-term addition.

The vacuum Alibaba is filling

The timing of this release is not accidental. The release arrives at a moment of remarkable upheaval in the AI video generation market. OpenAI discontinued Sora after it proved financially unsustainable. ByteDance indefinitely shelved the international rollout of Seedance 2.0 following a barrage of copyright complaints from Hollywood studios.

That leaves Seedance 2.0 technically at #1 on the leaderboard, but with constrained global availability. Seedance 2.0 powers the AI video features inside both CapCut and Dreamina, ByteDance's dedicated AI creative platform. The model targets text-to-video, image-to-video, and audio-conditioned video, with a companion variant called Seedance 2.0 Fast for lower-latency use cases. But for international developers who need reliable API access, ByteDance's copyright troubles create a real opening.

HappyHorse 1.1 also lands in a market where the global AI video generation space is growing fast. The global AI video generation market was valued at USD 554.9 million in 2023 and is projected to reach USD 2,172.4 million by 2032. Alibaba is positioning itself to capture a meaningful share of that growth, particularly in enterprise and creator workflows that need production-grade, multilingual output.

Who wins, who loses

For Alibaba, this is a validation of a strategy that started with a stealth launch and is now maturing into a competitive product. HappyHorse 1.1 is now live on Alibaba Cloud Model Studio with full API access for enterprise customers and developers, accompanied by a 40% sitewide launch discount for the first two weeks. The model is also available on fal.ai and Qwen Cloud, giving it broad third-party reach.

For developers choosing a video generation API, the leaderboard now looks like this:

RankModelCreatorEloPrice (per min, 1080p)
1Seedance 2.0 720pByteDance1,219$9.07
2HappyHorse 1.1Alibaba1,152$9.90
3HappyHorse 1.0Alibaba1,123$13.20
4SkyReels V4Skywork AI1,104$21.00
5Kling 3.0 1080p ProKlingAI1,104$20.16

The losers are the mid-tier models. For developers who need local inference, access to weights, or the ability to fine-tune, the existing open-weights leaders like Wan 2.2, HunyuanVideo-1.5, and LTX-2 remain the options. HappyHorse's advantage is production quality and speed via API, not open access. The gap between the top two models and everything else on the leaderboard is now substantial, which puts pressure on models like Kling and Google's Veo 3.1 to close the quality gap.

What this makes possible

The most practically significant capability in 1.1 is the nine-reference-image input. You can now feed up to nine reference images to anchor characters, environments, style palettes, or products across a multi-scene project. This is the production feature the creative community has been asking for: maintaining visual consistency across more than one or two clips without fine-tuning. For ad agencies and content studios, that is the difference between a demo and a real production workflow.

The multilingual lip-sync also opens up localization at scale. Phoneme-level lip-sync across seven languages, including English, Mandarin, Cantonese, Japanese, Korean, German, and French, is synchronized from frame one. Generating a product video that speaks natively in seven languages from a single prompt is a genuinely new capability for most production teams. Alibaba is not just competing on benchmark scores here. It is building a case that HappyHorse is the most practical choice for international content production at scale.