Summary of Key Points
Alibaba is simultaneously developing two video generation models, HappyHorse and Wan 3.0, which have highly overlapping functionalities. HappyHorse originated from Taotian Laboratory and focuses on commercial use cases, while Wan 3.0 is part of the Tongyi model ecosystem, emphasizing foundational capabilities. Although their initial development paths were different, as Wan 3.0 has evolved, their features (such as text-to-video and image-to-video generation, editing tools) and target user groups have begun to overlap, with both competing for market share on the same platform at varying prices. The teams responsible for these models have now been merged into the same business unit, yet they continue to update their respective models independently, leading to a waste of resources. The article argues that video generation models are capital-intensive and their business models are not yet stable. Alibaba should merge the two models as soon as possible to concentrate its resources on enhancing core capabilities rather than engaging in meaningless internal competition.
I. Internal Competition: From “Reasonable Attempts” to “Resource Waste”
Initially, it made sense for Alibaba to have the two teams follow different development paths:
- Wan is part of the Tongyi large-model ecosystem and functions more like a “universal toolkit,” focusing on technical fundamentals and developer ecosystems, suitable for foundational research.
- HappyHorse, developed by Taotian Laboratory, is designed for specific use cases such as advertising and short-form content creation, meeting user needs more directly.
However, the situation has changed:
- Functional Overlap: Wan 3.0 has incorporated many of HappyHorse’s core features (e.g., native audio, character consistency), and it is also cheaper (0.6 yuan per second in 720P resolution compared to HappyHorse’s 0.9 yuan).
- Team Merger: In June this year, the two teams were integrated into the Token Foundry business unit led by Wu Yongming, indicating that they should have been collaborating from the start.
- Customer Overlap: Both models compete for the same group of enterprises and creators on Alibaba’s Baolian platform.
Continuing with this internal competition is like having two similar washing machines at home: you need to buy separate detergents and fight over power outlets, which is not only costly and inefficient but also confusing for users.
II. The Video Model Field: Not Suitable for “Long-Term Competition”
Video generation models differ from traditional internet products due to two key factors:
1. High Resource Consumption: Training these models requires massive amounts of data and advanced computing power, making each iteration a significant investment.
2. Converging Technological Paths: The core capabilities of video models (duration, image quality, multi-modal input) are being pursued by all developers. HappyHorse and Wan 3.0 are now at similar levels; the differences are more about individual strengths (e.g., one may be better at character animation, another at shot composition), not a generational gap (like smartphones vs. feature phones).
It’s like two people climbing a mountain; if they start on different paths but realize the routes are converging halfway up, splitting again would be wasteful of effort. It’s more effective to merge their efforts and climb together.
III. The Illusory Advantages of All-Powerful Models and Low Prices
Neither Wan 3.0 nor HappyHorse possesses a core capability that users cannot do without:
- Wan 3.0 claims to generate 30-second videos and supports PPT input, but users want the content to be usable from the start. If the characters’ movements are stiff or the scenes are jerky, additional features will only increase the overall cost.
- HappyHorse is designed for advertising and short-form content, but if character consistency or product details (e.g., lipstick color) are inaccurate, businesses won’t use it.
- The Low-Price Trap: Wan 3.0’s lower price may seem attractive, but if it takes multiple attempts to produce a usable video, the total cost could be higher than HappyHorse’s. Users value cost-effectiveness, not just low prices.
For example, when choosing a barber, you would prefer one who is skilled at cutting short hair quickly and well over one who can do everything but consistently produces inconsistent results. The same applies to video models; being all-purpose or offering low prices without specific strengths is meaningless.
IV. The Optimal Solution: Unified Foundation with Differentiated Applications
The article suggests the following approach for Alibaba:
- Unified Foundation: Merge Wan 3.0 as the base model and incorporate HappyHorse’s expertise in commercial applications (advertising, short-form content) and character performance to enhance core capabilities (e.g., character consistency, smooth animations).
- Branded Differentiation: Keep HappyHose as a dedicated product brand for creators, providing templates and workflows (e.g., one-click ad script generation) to simplify user experiences.
This approach avoids reinventing the wheel and ensures both technical depth and user satisfaction, similar to how smartphone manufacturers use a unified chip platform for various models (e.g., Huawei’s Mate and P series).
V. Alibaba’s “Mindset Inertia”: Past Approaches from the Pre-AI Era
Alibaba’s past success with internet products (e.g., e-commerce, social media) relied on internal competition to quickly test ideas, with the winner taking all the benefits. However, AI models are different:
- Testing ideas is much more expensive for AI (training a model can cost millions).
- Internet products succeed through rapid feature updates, while AI models require breakthroughs in core technologies.
Alibaba needs to shift from an “internet product mindset” to an “AI model mindset.” Continuing with internal competition will only scatter resources and prevent it from establishing a competitive advantage in the video generation field.
In Conclusion
Rather than having the two models compete for customers, Alibaba should combine their strengths to focus on developing unique capabilities that no other company offers. This is the key to success in the AI era.