Physix Frontier70% Cost, 100% Anxiety: What Is Open-Source Video Generation Really Competing On?
Last week in Shenzhen, I checked out a robotics project. The founder showed me a demo: a girl with long pink hair posing in front of the camera, smooth movements, natural lighting and shadows. He said it was run using an open-source model, costing less than two mao (approx. $0.03 USD). When I asked what he used, he said MiniMax's newly open-sourced HunyuanVideo-7B, which is 70% cheaper than the previous Seedance.
I immediately pulled out my phone to check. Seedance is ByteDance's closed-source model, while HunyuanVideo is MiniMax's open-source version. Saying it's 30% off is understating it; calculated by token consumption, it might be even lower. Where is the ceiling for this track? I review 500 projects a year. Video generation has been a hot topic since last year, but the consensus before was that models capable of doing it well had extremely high costs, making them unaffordable for small teams. Now that open-source models have dropped prices by 70%, it's like smashing the barrier from "tens of millions" down to "millions."
1. Valuation Logic Has Changed: From "Buying Models" to "Buying Ecosystems"
The valuation logic for video generation models used to be "technical leadership + closed-source charging." Seedance's pricing strategy is a typical B2B approach: charging per minute, with enterprise clients paying tens to hundreds of yuan per video segment. The problem with this model is that the ceiling is visible. Based on current total domestic demand for video generation, even if penetration reaches 30%, the market size is only around ten billion yuan. Moreover, technology iterates quickly; leadership from three months ago can be caught up.
MiniMax's approach is to open-source their self-developed models while providing lower-cost API services. This is behind a classic "dimensional reduction attack": using open-source models to disrupt market pricing, then earning money through ecosystem services and computing power. The business logic of open source has never been "free," but rather "low-cost customer acquisition + high-value conversion." Just like how Android was free back in the day, Google made money through ads and GMS. MiniMax's computing pool and cloud services are the real sources of profit.
I did some math: Suppose a developer uses HunyuanVideo-7B to generate 1,000 videos, saving 700 yuan compared to Seedance. If this developer generates 100,000 videos a month, the savings are enough to buy an A100 GPU. But MiniMax can recoup 300 yuan of those 700 yuan by offering tuning services, model hosting, and data annotation. In the long run, the developer's migration costs will become increasingly higher.
2. Business Model Trade-offs: Is Open Source a "Traffic Driver" or the "Main Course"?
Many founders ask me: Will open-sourcing starve us to death? The answer is: It depends on how you do it. MiniMax's layout is clear: open-source models are the "traffic drivers," while the true "main course" is enterprise-level services. Listed on their official website, the "Enterprise Edition" includes: private deployment, customized training, and 7x24 technical support. These services have very high unit prices, and once customers use them, replacement costs are huge.
But there's a risk here: The quality of open-source models must be good enough to attract developers. Otherwise, people would rather use Seedance, even if it's more expensive, for stability. MiniMax's HunyuanVideo-7B scores close to Seedance in authoritative benchmarks on certain metrics (like motion continuity, light/shadow consistency), but its inference speed is significantly slower. A 70% price discount for 70% speed—this trade-off may not be worth it for developers.
More critically, the technical barriers for video generation models are lowering. Last year, everyone was competing on "generation quality"; this year, competition has shifted to "controllability" and "interactivity." MiniMax's open-source strategy is essentially betting that technology will rapidly commoditize, securing the ecosystem position first. If open-source models become the standard for the developer community, subsequent value-added services will follow naturally.
3. Where Are the Competitive Moats? Computing Power, Data, or Trust?
I believe there are three layers to the competitive moats for video generation models:
1. Computing Cost: MiniMax is backed by a major tech giant, ensuring its computing pool. However, open-source models mean anyone can deploy them themselves, diluting the computing advantage over time. Unless MiniMax maintains continuous leadership in model compression and inference optimization, this moat won't hold for a year.
2. Data Flywheel: Open-source models attract more developers, and the data generated by these developers can feed back into improving the models. But this assumes developers are willing to share data, or that MiniMax can collect high-quality data via APIs. Currently, MiniMax's API pricing is very low, so developers might prefer using the API directly rather than deploying open-source models themselves—allowing MiniMax to capture the data.
3. Developer Trust: The biggest benefit of open source is transparency. Developers can inspect code and modify models without worrying about being "harvested" by closed-source companies. But trust takes time to build, and MiniMax's community operation capabilities are a weak point. They don't have the robust community mechanisms of HuggingFace; many developers might just try it out and leave.
Investment Judgment: In the short term, MiniMax's open-source strategy can quickly grab market share, but in the long term, we need to see if they can convert "traffic" into "retention." If, in three months, 80% of video generation projects in the developer community are based on HunyuanVideo-7B, then this investment is worthwhile. If it's just a flash in the pan, even a 70% discount won't save it.
My personal opinion is: This track is currently too crowded, but open-source models give small teams a way to survive. If I were an entrepreneur, I would use open-source models for vertical scenario applications, such as virtual streamers, e-commerce short videos, and educational training. These scenarios don't require the absolute top-tier models but need rapid iteration and low costs. MiniMax's open-source models fit exactly into this niche.
One-line summary: A 70% discounted model isn't the end point, but the starting point of ecosystem warfare. Whoever gets developers addicted first wins.