Video generation has developed a split that image generation went through two years ago. On one side sit closed models reachable only through a vendor’s endpoint. On the other sit open-weight families — Alibaba’s Wan series being the most prominent — that can be downloaded, self-hosted, fine-tuned, and also consumed through hosted endpoints when that is easier.
For a technical team choosing where to build, that distinction matters more than the benchmark numbers everyone quotes.
What Open Weights Actually Buy You
Three things, and it is worth being precise about them because vendors on both sides overstate the case.
- No dependency on one company’s uptime or pricing. If a hosted provider raises rates, deprecates a version or has an outage, an open-weight model can be moved to different infrastructure. That optionality has real value for a product whose core feature depends on generation.
- Data residency and privacy control. If prompts or source images cannot leave your infrastructure — common in regulated industries — self-hosting is the only option, regardless of what quality is available elsewhere.
- Fine-tuning on your own material. Adapting a model to a specific visual style, product range or character set is possible when the weights are available and generally is not otherwise.
What open weights do not buy you is cheapness by default. Which brings us to the part most teams get wrong.
Self-Hosting Is Rarely Cheaper Below Serious Volume
The arithmetic is unforgiving. Video generation requires substantial GPU memory and produces output over tens of seconds per clip. A GPU instance capable of running a current-generation video model costs money every hour it exists, whether you generate one clip that day or four hundred.
Hosted access, by contrast, is billed per second of generated video. Rates for the Wan 3.0 API and comparable models are published openly on aggregation platforms that expose several video models through one account, so the crossover point can be calculated rather than guessed.
Work it out before committing to infrastructure. Take your realistic monthly clip count, multiply by average clip length, multiply by the published per-second rate, and compare against the monthly cost of an instance that stays warm enough to serve requests without a cold-start delay your users will notice. Most teams discover the crossover sits far higher than they assumed, and that hosted access is the correct answer until the product has real traction.
The Hybrid Pattern That Usually Wins
Teams that end up happy generally do not choose one side permanently. They start on hosted endpoints because it removes infrastructure work entirely, keeps the code provider-agnostic, and lets them compare several models on their actual workload rather than on demo reels.
They keep an eye on two triggers: sustained volume that makes a dedicated instance cheaper, and a requirement — privacy, fine-tuning, or a model version the vendor has retired — that hosted access cannot satisfy. When either trigger fires, having chosen an open-weight family from the start means the migration is an infrastructure project rather than a rewrite.
That is the strongest practical argument for open-weight models even when you have no intention of self-hosting today: it keeps the door open at no cost.
Engineering Details That Apply Either Way
Treat generation as an asynchronous job from the beginning. Clips take tens of seconds, users background apps, and a synchronous request will be reported as a bug in the first week.
Expect a meaningful reject rate. Motion introduces failure modes that still images do not have — deformation part-way through a clip, drifting camera moves, physics that look wrong — so generate short and extend only what works.
Keep the prompt, the seed and the model version alongside every stored output. When someone asks six months later how a clip was produced, or when you need to reproduce a look after a version change, that record is the only thing that helps.
Abstract the provider behind a thin internal interface. This costs an afternoon and is what makes the eventual migration, in either direction, a small piece of work.
Evaluating a Video Model on Your Own Workload
Benchmark reels are produced by people who know exactly which prompts the model handles well. They tell you almost nothing about your use case. Run your own comparison instead, and keep it small enough to finish in a day.
Collect eight shot descriptions from work you actually need: an establishing shot, a product in motion, a character action, a camera move, a texture close-up, something with water or fabric, something with a crowd, and one deliberately awkward request. Generate three clips per description per model at the same settings.
Score four things. Did it follow the description, particularly camera movement and timing? Does motion hold together for the full duration without deformation? Is the output usable without re-rendering? And when it fails, does it fail mildly or catastrophically?
Then compute cost per usable clip rather than cost per second. A model with a lower rate and a 30 percent hit rate is more expensive than one costing twice as much with an 80 percent hit rate, and that ordering is invisible in any published comparison.
Where Video Models Still Struggle
Being specific here saves wasted generations. Legible text rendered inside a clip remains unreliable across every model available, so titles and captions belong in your editing layer. Maintaining the same character or the same room across a sequence of clips requires additional machinery rather than prompting alone. Precise physical simulation — liquids pouring correctly, cloth settling naturally, hands manipulating objects — is where most rejected clips fail. And anything depicting a real person, product or location comes out approximate, which matters commercially as much as technically.
The Honest Summary
Open weights are about control and optionality, not about saving money at small scale. Hosted endpoints are about shipping quickly without an infrastructure team. Choosing an open-weight family accessed through a hosted endpoint gets most of both, and defers the expensive decision until you have the volume data to make it properly.
