Open Source vs. Closed AI Video Models, Explained

We run several video models side by side — some open-weight, some closed and API-only — and users mostly just pick a model name from a list without seeing how differently each one got there. Wan 2.2, for example, is released on GitHub under the Apache 2.0 license with downloadable weights; Kling and Seedance are closed source, with no public model repository and access controlled entirely through the vendor's own API. That split shapes cost, reliability, and how much control a platform has over what it's running — worth explaining plainly instead of treating every model on a list as interchangeable.

Compare models and generate

What 'open source' actually means for a video model

Wan 2.2 was released by Alibaba's Wan team under the Apache 2.0 license, with model code and weights published on GitHub (Wan-Video/Wan2.2) and mirrored on Hugging Face and ModelScope. Apache 2.0 is a permissive license: it allows commercial use, modification, and redistribution as long as the original copyright notice is retained. In practice that means anyone with the hardware can download the weights and run the model themselves — the smaller 5B TI2V variant is built to run on a single consumer GPU like an RTX 4090, while the larger 27B-parameter Mixture-of-Experts models (14B active per step) need real infrastructure to serve at scale. Either way, there's no vendor API sitting between you and the model.

Closed models: capable, but access runs through the vendor's door

Kling and Seedance work differently. Both are closed source — no published weights, no model repository, no way to inspect or run either one outside the vendor's own platform. Kling's API has a real barrier to entry: having a funded Kling account doesn't automatically grant API access, you have to separately subscribe to an API plan through the developer console, and the official route often involves enterprise approval or a meaningful monthly commitment (which is why third-party API aggregators exist at all). Seedance, from ByteDance, is proprietary in the same way — there's no official GitHub release of weights, and access runs through BytePlus/Volcano Engine's Ark API. That access isn't even guaranteed to stay open: ByteDance suspended Seedance 2.0's overseas API in March 2026 over a copyright dispute, then reopened it for enterprises and developers through BytePlus the following month.

The quality gap has narrowed — that's not the whole trade-off

Open-weight models have closed a lot of the gap with closed competitors on motion consistency and visual fidelity, and running one locally means no per-generation API bill and no dependency on a vendor keeping an endpoint online. But an Apache 2.0 license doesn't come bundled with hosting or moderation — running a 27-billion-parameter MoE model at real scale is genuine infrastructure work, and any content-safety layer has to be built by whoever deploys it, not inherited from the license. Closed models push that infrastructure burden onto the vendor, but in exchange your capability is whatever that vendor's API currently offers, at whatever price and access tier they've set — and as Seedance's own access suspension showed, that can change without much warning.

Why we run both kinds side by side

Our video lineup mixes the open-weight Wan alongside closed, API-based models like Kling and Seedance, plus Vidu and Hailuo — not because one category is objectively better, but because they solve different problems. Open weights give us direct control over exactly what a job is running and how it's hosted; closed models give access to capability at a scale we couldn't train or serve ourselves. Picking a model for a given generation is as much a licensing and infrastructure decision as a creative one, even though the person clicking 'generate' mostly just sees the output.

What the license does and doesn't cover

Apache 2.0 covers the code and model weights — it isn't a blanket rule about what the output is allowed to be used for. Wan's own usage terms still prohibit using the model to generate content that violates the law, causes harm, or targets vulnerable people, regardless of how permissive the underlying code license is. 'Open source' describes who can run the model and how; the responsibility for what gets generated with it still sits with whoever deploys it, open or closed.

FAQ

Is Wan 2.2 actually free to use commercially?

Yes — it's released under the Apache 2.0 license, which permits commercial use, modification, and redistribution as long as the original copyright notice is retained. That's separate from Wan's own usage terms, which still prohibit generating unlawful or harmful content regardless of the license.

Can I download Kling or Seedance and run them on my own hardware?

No. Both are closed source — there's no published model repository or downloadable weights for either. Access is vendor-hosted only: Kling through its developer console API plans, Seedance through ByteDance's BytePlus/Volcano Engine Ark platform.

Does open source mean lower quality?

Not reliably anymore. Open-weight models like Wan 2.2 have closed much of the gap with closed competitors on motion consistency and visual fidelity, though how any two specific models compare still depends on the exact task and prompt.

Why bother with a closed, API-only model instead of just using open weights?

Because closed vendors are training and hosting models at a scale most platforms couldn't build or run themselves. The trade is giving up local control and becoming dependent on that vendor's API access and pricing, which isn't guaranteed to stay stable — Seedance's own overseas access suspension in March 2026 is a real example of that risk.

Which model does Uncutly actually use?

We run several video models side by side, including the open-weight Wan alongside closed API models like Kling and Seedance. Which one handles a given generation depends on the model you pick — there's no single default engine.