Alibaba has built a two-tier video generation system that operates as fundamentally different products despite coming from the same company. HappyHorse 1.1, which debuted anonymously and won blind-test leaderboards in April 2026, is the closed flagship model optimized for audio-native video generation. Wan 2.7, launched earlier in April, is the cheaper, more editable workhorse with open-source weights. Understanding them as a stack rather than competing products reveals how Alibaba is positioning itself as OpenAI's Sora and ByteDance's Seedance retreat from the market.

How Did an Anonymous Model Win the Video Generation Race?

HappyHorse's origin story is unusual in the 2026 AI landscape. Around April 7, 2026, an unbranded model calling itself "HappyHorse" appeared on the Artificial Analysis Video Arena, a blind-test platform where voters compare video clips without knowing which model produced them. The model immediately took the number one spot on both text-to-video and image-to-video leaderboards, ahead of established competitors including ByteDance's Seedance line. Three days later, on April 10, Alibaba confirmed authorship.

This anonymous-debut strategy was deliberate. By winning a blind test before revealing the brand, Alibaba neutralized a discount that Western evaluators sometimes apply to Chinese-lab releases. Voters ranked the output first and learned the vendor afterward, a reversal of how Alibaba launched Qwen-Image-3.0 three months later with maximum benchmark transparency. That inconsistency signals Alibaba is still experimenting with how to position different products in the market.

What Are the Key Differences Between HappyHorse and Wan?

The two models serve different production needs and price points. HappyHorse 1.1 shipped June 21-22, 2026, while Wan 2.7 launched April 6 with incremental updates through June 12. Neither is new as of late July, which is why understanding them as a stack decision matters more than chasing release announcements.

• HappyHorse 1.1 (Flagship): Unified audio-video generation in a single pass, native lip-sync across seven languages (Mandarin, Cantonese, English, Japanese, Korean, German, French), four generation modes including video editing, 1080p output, 3-15 second clips, and five aspect ratios. No open weights, no official technical report. Costs approximately $0.14 per second at 720p and $0.28 per second at 1080p on fal.ai.
• Wan 2.7 (Workhorse): Apache 2.0 open-source license for base models, 27 billion total parameters with 14 billion active mixture-of-experts architecture, seven generation modes including natural-language video editing, published architecture documentation. Costs approximately $0.10 per second on fal.ai, making it the cheaper option on the same hosting platform.
• Architecture Difference: HappyHorse uses a unified 40-layer self-attention Transformer with no cross-attention modules, generating audio and video jointly in a single pass. Wan 2.7 is a modular suite with four sub-models, offering more flexibility for teams that want to customize individual components.

Pricing varies by hosting platform. On fal.ai, Wan 2.7 runs $0.10 per second flat while HappyHorse runs $0.14 per second at 720p and $0.28 per second at 1080p. Hedra's July comparison lists different per-minute rates on its platform. There is no single canonical price across all providers.

Why Does Alibaba's Openness Strategy Seem Inconsistent?

Alibaba ships three different openness postures across three modalities, a pattern that suggests either deliberate product segmentation or organizational misalignment. Wan ships Apache 2.0 weights and published architecture, making it accessible to researchers and developers who want to modify or fine-tune the base models. HappyHorse ships neither weights nor a technical report, keeping the flagship closed. Qwen-Image-3.0, Alibaba's image model, went invite-only with no published benchmarks.

For production teams, this inconsistency creates a practical choice: use the open, cheaper, more customizable Wan 2.7 for internal workflows and experimentation, or pay more for HappyHorse's closed, optimized flagship performance when output quality is the priority. The two-tier approach mirrors how other AI companies segment their offerings, but Alibaba's lack of transparency about why each model has a different openness posture makes the strategy harder to predict or plan around.

How Do Arena Leaderboard Rankings Actually Compare?

HappyHorse's competitive standing has drifted since its April debut, and the published arena snapshots do not reconcile with each other. VentureBeat, citing Arena.ai on June 22, reported HappyHorse at an Elo rating of 1,444 and number two across leaderboards. A July 17 Hedra read lists 1,149 for text-to-video and 1,111 for image-to-video. Different platforms, different dates, and different voter pools mean these numbers should never be merged into a single figure.

The lesson for evaluating video models is that leaderboard position is a snapshot, not a permanent ranking. Voter pools change, compression levels vary, and different platforms weight different aspects of quality. A model that ranks first on one platform on one date may rank second on another platform a month later, not because the model degraded but because the evaluation context shifted.

What Does This Mean for the Broader Video Generation Market?

The competitive field around Alibaba has contracted significantly. OpenAI discontinued its standalone Sora product, and ByteDance shelved its international Seedance rollout, leaving Alibaba, Google, xAI, and Black Forest Labs as the primary contenders in the text-to-video space. Alibaba's two-tier stack strategy positions it to capture both the premium quality segment (HappyHorse) and the cost-conscious, customization-focused segment (Wan 2.7) simultaneously.

For creative production teams, the practical implication is that Alibaba is now offering a choice that competitors do not: a closed, optimized flagship for when output quality is non-negotiable, and an open, cheaper alternative for when flexibility and cost matter more. The fact that both models are available on the same hosting platforms like fal.ai means teams can test both and choose based on their specific workflow, rather than being locked into a single vendor's approach.