Open Weights DeskWhich video models you can actually download, which you cannot, and what the difference costs.

Wan 2.2 is the last flagship in this family with published weights

If you searched for "Wan 3.0 open source" and landed somewhere that implied you could download it, here is the actual state of the family, with the line drawn in the right place.

Generation	Weights published	Licence
Wan 2.1 (wanx2.1-t2v-*)	Yes	Open
Wan 2.2 (July 2025)	Yes	Apache 2.0
Wan 2.5	No	—
Wan 2.6	No	—
Wan 2.7	No	—
Wan 3.0	No	—

Checked 27 August 2026.

Wan 2.2 is the boundary. It shipped in July 2025 under Apache 2.0 with a real set of checkpoints — Wan2.2-T2V-A14B, I2V-A14B, TI2V-5B, S2V-14B, Animate-14B — and the licence is genuinely permissive: no revenue ceiling, no field-of-use restriction, and the publishers explicitly disclaim rights over what you generate with it. The 5B variant runs in about 8GB of VRAM once quantised, which puts it inside a consumer card.

Everything after it is closed. Wan 2.5, 2.6, 2.7 and 3.0 have no public repository and no weights on any model hub. There is no ComfyUI node for Wan 3.0 because there is nothing for a node to load.

The inference that keeps being wrong

The reasoning that produces bad advice goes: Wan is an open-source family → Wan 3.0 is a Wan model → therefore Wan 3.0 can be self-hosted.

The first premise was true and stopped being true a generation and a half ago. The family forked into an open line that ended at 2.2 and a closed line that continues. Both lines carry the same brand, which is why the confusion is durable rather than careless.

There is a reliable tell for reading any product page in this space: if a site claims to offer Wan 2.5 or later and also implies it runs its own inference, one of those two claims is false. There are no weights to run. It is calling an API, which is fine, but it is a different thing and it has different implications for latency, privacy and what happens when the upstream changes.

What the licence buys, and what it does not

Apache 2.0 on Wan 2.2 removes the licence fee. It does not remove the GPU bill, and for video generation the GPU bill is essentially the whole cost.

That is worth being concrete about, because "free and open source" gets read as "free", and the second word is doing no work at all in this sentence:

• Apache 2.0 frees the weights. You can download, modify, redistribute, and use commercially, with attribution, and you own your outputs.
• It does not free the hardware. A 14B video model wants a serious card, and it wants it for the duration of every generation. A 5B variant fits in 8GB but is a smaller model, which is a quality decision, not a free lunch.
• It does not free the operations. Weights are a file. A service is a queue, a scheduler, a storage layer, a moderation pass and someone awake when it breaks.

And it does not backfill capability. Wan 2.2 does not have the things people actually want from 3.0: no thirty-second single pass, no 480P draft tier, no document input, no web page input, no video editing pass. It is a good model from July 2025 and it is a generation-and-a-half behind what it is being mistaken for.

When the open model is genuinely the right answer

There are real cases, and they are not about price:

• The weights cannot leave your premises. Regulatory, contractual, or the material is simply not going to a third-party API. This is the strongest case and price is irrelevant to it.
• You want to fine-tune. You cannot fine-tune a closed endpoint. If adapting the model to a specific look or subject is the point, an open model is the only kind that works.
• You need offline or air-gapped operation.
• You have idle GPUs already paid for. Marginal cost changes the arithmetic completely.

Notice that none of those is "it is cheaper", and the next note works through why.

If none of those apply to you and the reason you were looking for open weights was cost, the honest recommendation is to stop looking for them and go and read a per-second price list instead — including what one clip on the current closed model actually costs, which is the number the self-hosting plan has to beat.

If you do need weights, name the requirement first

Two questions settle it faster than any comparison table:

1. Does the material have to stay on hardware you control? If yes, you are choosing between open models and nothing else, and the current-generation closed models are not in the conversation regardless of what they can do.
2. Do you need to change the model's behaviour, not just its prompt? If yes, same answer.

If both are no, weights are not what you were looking for. What you were looking for was a lower bill, and that is a different search.