NodeValid

Published video-model weights, licences and hardware notes

MAGI-1 ships 24B and 4.5B with own notes

Two sets of weights are published, MAGI-1-24B and MAGI-1-4.5B, and each carries its own hardware note. One names eight H100 or H800 cards; the other names one RTX 4090. As of 2026-09-22.

MAGI-1 in the size column, as publishedPublishing two sizes with two separate hardware statements makes the count the index for everything else the repository says about running it.Fixed by the published countLeft to the runThe count24B and 4.5B models—What it belongs toBoth published, each with its ownhardware note—Licence on that releaseApache License 2.0—Memory for a run—H100 or H800 x 8 for the 24BOutput size—720 x 720 as the 4.5B defaultThe count selects the machine
Fig. 1 One repository, two very different machines, and the count telling you which is which.
MAGI-1 in the size column, with the columns that qualify it. Read from github.com/SandAI-org on 2026-09-22.
ColumnWhat the pages state
Size, in the vendor's own words24B and 4.5B models
The detail published beside itBoth published, each with its own hardware note
Licence, from the same pagesApache License 2.0
Hardware, from the same pagesH100 or H800 x 8 for the 24B

Inclusion rule. The value this page is about, the detail the vendor attached to it, and the two columns that change how it should be read. A column the vendor left empty keeps its row and says so. Order. This column first, then its detail, then the two columns that qualify it.

1Two audiences in one release

Eight data-centre cards and one consumer card are not two points on a scale; they are two different kinds of user. Publishing both sizes with both notes says the vendor knew that.

It also makes the repository unusually easy to act on. A reader knows within one line whether the release is for them, without reading a paper or a benchmark.

2The smaller size has a floor and a lower one

For the 4.5B the repository states at least 24GB of GPU memory, and adds that a distilled fp8 configuration works with at least 12GB. The count stays the same while the requirement halves.

That is a build being documented rather than a model, and it is the clearest example in the register of a memory figure describing software choices instead of weights.

3What the count does not decide here

The default resolution of 720x720 is described as set for speed, with arbitrary resolutions supported, so the output size is a configuration rather than a property of either size.

Length is not bounded at all: frames are an argument with no stated maximum. The size column tells a reader which machine; the rest of the release is still open.

4Sources

The count comes from github.com/SandAI-org, read 2026-09-22, exactly as the vendor wrote it. Everything else this release publishes is on its family entry; why the count sits in its own column is on size. Nearby: Counted to the unit, The 5B that runs at home.