Step-Video-T2V: what the repository publishes

Step-Video-T2V publishes 30 billion parameters under MIT, the shortest licence text in this register attached to its largest published model. Peak GPU memory is stated per configuration: 78.55 GB at 768px768px with 204 frames, 72.48 GB at 544px992px with 136. As of 2026-09-22.

Inclusion rule. The same five fields for every family here. A field the repository does not state is recorded as not stated rather than estimated from the parameter count. Order. Fixed field order, identical on every family page.

1A four-paragraph licence on thirty billion parameters

MIT imposes attribution and stops. It is the shortest licence in this register and it sits on the largest model here, which is a clean demonstration that the licence column and the hardware column have nothing to say to each other. Permission does not predict cost.

Read practically, the rights question closes in a minute and the capacity question stays wide open. A permissive licence on a model that peaks near 79 gigabytes is permission to rent a card, not permission to avoid renting one.

2Three configurations, three figures, and the frame count is the dial that moves them

The repository publishes a table rather than a number: 78.55 GB at 768px768px with 204 frames, 77.64 GB at 544px992px with the same frame count, and 72.48 GB at 544px992px with 136 frames. Cutting 68 frames saves about five gigabytes. Changing the shape of the frame saves about one.

This is the rare entry where a reader can see which dial matters. Length moves this model's memory several times as much as resolution does, which inverts the assumption most capacity planning starts from, and it is only visible because three rows were published instead of one.

3What 204 frames is, and what the repository declines to convert it into

The stated ceiling is expressed in frames: up to 204 of them. No frame rate travels with it in the requirements table, so how many seconds that amounts to depends on a number the reader has to supply, and this register does not supply it for them.

The memory figures carry their own qualifiers: batch size one, without cfg distillation, with multi-GPU deployment discussed separately. Those clauses are the difference between a figure somebody can reproduce and a figure that merely sounds like a specification, so the row keeps them attached.

4The same field, across every family

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

Inclusion rule. Families whose vendor has published weights openly. A family whose repository does not state this particular field keeps its row, marked hollow. Order. Alphabetical by family name.

• Licence
  mit, the same tag on the repository and on the model cardstated in the repositorystepfun-ai, Step-Video-T2V repository / recorded 2026-09-22
• Model size
  30 billion parametersfrom the repository's own introductionstepfun-ai, Step-Video-T2V repository / recorded 2026-09-22
• Hardware
  78.55 GB peak GPU memory at 768px768px204f, 77.64 GB at 544px992px204f and 72.48 GB at 544px992px136fstated per configuration at batch size onestepfun-ai, Step-Video-T2V repository / recorded 2026-09-22
• Length
  Videos up to 204 framesin frames, with no frame rate attachedstepfun-ai, Step-Video-T2V repository / recorded 2026-09-22

5Sources

Read from github.com/stepfun-ai on 2026-09-22. Hardware figures across every family are collected on requirements; the licence column is set out on licence and what counts as a published weight is on how read. Previous family: Allegro. Next family: Open-Sora.