Open weights versus closed API: renting or owning the model

Think of it as renting glass versus owning it. A rental house has the newest lenses and services them for you, but it can swap the stock, retire a set, or change what it will hand over, and the lens you shot your test on in March may not be on the shelf in September. Glass you own is yours: older, harder to look after, and identical in three years. A closed generation API is the rental house. You get the leading quality, someone else's GPUs, safety tooling, and in some cases contractual indemnity, but the endpoint can be revised or deprecated between your test and your delivery, and you cannot pin a version you do not possess. Open weights are the owned lens. You download a file, keep it on your own hardware, run it offline on material that cannot leave the building, fine-tune it on your own character or look, and reproduce a shot next year because nothing moved. What you take on is GPU capacity, engineering time and maintenance. The important trap is vocabulary. Open weights is not open source. The Open Source Initiative's Open Source AI Definition asks for information about the training data, the training and inference code, and the weights; most released video models do not meet that bar. The licences vary widely and they are real contract terms. Wan and Mochi 1 are released under Apache 2.0. CogVideoX is split, with the 2B model under Apache 2.0 and the larger models under a bespoke licence. HunyuanVideo ships under the Tencent Hunyuan Community License, whose text states it does not apply in the European Union, the United Kingdom and South Korea. Stable Video Diffusion sits under a Stability community licence that routes commercial use through separate terms. This is a pipeline trade-off, not a moral one.

What people get wrong

Each of these is a belief worth dropping, followed by what is actually happening.

• open weights means open source. The Open Source Initiative's definition requires information about the training data, the training and inference code, and the weights. Most downloadable video models ship under custom licences that do not satisfy it.
• open means free to use commercially anywhere. HunyuanVideo's community licence states it does not apply in the European Union, the United Kingdom or South Korea, and Stable Video Diffusion's community licence directs commercial use to separate Stability terms. Licences differ per model and sometimes per model size.
• the closed model is always better. Hosted flagships often lead on quality, but the gap moves with every release, and quality is one axis alongside control, confidentiality, reproducibility and cost.
• pinning the model name in an API call gives identical results forever. Hosted endpoints are updated underneath the name. Only weights you hold can be genuinely pinned.
• self-hosting is cheaper. You swap a per-generation charge for GPUs, storage, engineering time and idle capacity. It pays off at volume and on confidential work, not automatically.

Terms this introduces

Open weights

A model whose trained parameters can be downloaded and run locally, regardless of whether its licence meets any open-source standard.

Open Source AI Definition

The Open Source Initiative's 1.0 standard, which asks for information about training data, the training and inference code, and the weights.

Version pinning

Locking a production to one exact model build so results stay stable across a revision cycle.

On-premises inference

Running generation on hardware you control so confidential material never leaves your network.

Community licence

A vendor-written licence that permits broad use with carve-outs, such as territorial exclusions or separate commercial terms, and is not an open-source licence.