The version sold is not the version published

The model vendor's own organisations publish weights for 2.1 and 2.2. Version 3.0, which is what the hosted service here sells by the second, has no published weights. That gap is the central finding of this register. As of 2026-09-12.

Inclusion rule. Weights published by the model vendor's own organisation. Community re-uploads and derivatives are not counted. Order. By version number.

1An open family is not an open version

Reading that a family is open source and concluding that the current generation can be run locally is the most common error this register exists to correct. Two generations are published; the one being sold is not, and nothing about the first fact implies anything about the second.

A benchmark run on published weights therefore describes an earlier generation. Quoting it as evidence about what an API delivers compares two different models that share a name.

2The lag is the thing to watch, not the fact of it

Publishing a generation behind what you sell is an ordinary commercial arrangement and not a criticism. What matters to a planner is how far behind and whether the distance is growing, and that is answerable only by watching the repository over time.

This register records what is published today with a date on it. Two readings months apart are what turn a single row into a trend, and the dates are there so that comparison is possible.

3What a team can actually do with 2.1 and 2.2

Run them, modify them, ship products on them, with no licence obstacle. What they cannot do is reproduce what the hosted 3.0 tiers produce, and no amount of local tuning closes a generation gap.

For research, for cost control on undemanding work, and for anything that must keep running without a vendor, the published generations are a real option. For matching the output of a paid tier, they are not the same thing.

• Published versions
  The vendor's own organisations publish weights for 2.1 and 2.2no 3.0 listedWan-AI on Hugging Face / recorded 2026-09-12
• Version 3.0
  No weights published (as of 2026-09-12)the one the hosted service here sellsWan-AI on Hugging Face / recorded 2026-09-12

4Open weights, hosted

Instead of local hardware for “Which version”, SceneMixer runs hosted models; its guides describe the result:

The longest an episode may run: 240 seconds for Short Drama, 300 for TV Series, none for Film.

SceneMixer, Short Drama, TV Series or Film — “What is the episode cap?” (read 2026-10-13)

Paste a scene or a short script and SceneMixer gives you a storyboard — numbered shots with framing and seconds — and then a moving version of each segment, which you can join into a rough cut of the scene.

SceneMixer, Using SceneMixer for Previs and Animatics Before a Live-Action Shoot (read 2026-10-13)

5Sources

Repository statements and hosted rates are linked from the model pages and the method page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Running it locally, 4090 or 80GB, Clip length.