NodeValid

Published video-model weights, licences and hardware notes

Counting cards against counting gigabytes

MAGI-1 states H100 or H800 x 8 for its 24B model and one RTX 4090 for the 4.5B. Open-Sora states 60.3GB peak on one GPU and 44.3GB across four or more, on H100 or H800. Two units for one question. As of 2026-09-22.

MAGI-1 and Open-Sora on the same columnA card count carries memory, numeric formats and interconnect together; a per-card gigabyte figure carries only memory and has to name the cards separately.MAGI-1Open-SoraHardware, as each vendor statesitH100 or H800 x 8 for the 24B60.3GB peak at 768px on one GPUThe detail published beside itOne RTX 4090 for the 4.5B; 12GB for adistilled fp8 build44.3GB across four or more, on H100 orH800Resolution, from the same pages720 x 720 as the 4.5B default256px and 768pxNeither unit converts into the other
Fig. 1 Two units for describing a multi-card run, each losing something the other keeps.
What the repositories actually stateFilled where the repository states something, hollow where it does not. A blank is not an inference.What the repositories actually stateMAGI-1Open-SoraHardware, as each vendor states…Hardware, as each vendor states itHardware, as each vendor states it — MAGI-1: H100 or H800 x 8 for the 24BHardware, as each vendor states it — Open-Sora: 60.3GB peak at 768px on one GPUThe detail published beside itThe detail published beside it — MAGI-1: One RTX 4090 for the 4.5B; 12GB for a distilled fp8 buildThe detail published beside it — Open-Sora: 44.3GB across four or more, on H100 or H800Resolution, from the same pagesResolution, from the same pages — MAGI-1: 720 x 720 as the 4.5B defaultResolution, from the same pages — Open-Sora: 256px and 768pxLength, from the same pagesLength, from the same pages — MAGI-1: No ceiling recordedLength, from the same pages — Open-Sora: num_frames of 4k+1, under 129
Fig. 2 Filled where the repository states something, hollow where it does not. A blank is not an inference.
MAGI-1 and Open-Sora on the hardware column, read from github.com/SandAI-org and github.com/hpcaitech. Recorded 2026-09-22.
PointMAGI-1Open-Sora
Hardware, as each vendor states itH100 or H800 x 8 for the 24B60.3GB peak at 768px on one GPU
The detail published beside itOne RTX 4090 for the 4.5B; 12GB for a distilled fp8 build44.3GB across four or more, on H100 or H800
Resolution, from the same pages720 x 720 as the 4.5B default256px and 768px
Length, from the same pagesNo ceiling recordednum_frames of 4k+1, under 129

Inclusion rule. Two families whose published values in this column lead to opposite decisions. A pair that happens to agree is not listed, because agreement is already visible on the column summary. Order. The column itself first, then its detail, then the columns that qualify it.

1Eight cards is not a memory total

Read as gigabytes, eight H100s is a large number and a useless one. What the count actually specifies is a node: the interconnect, the driver stack and the formats the run was exercised against.

Open-Sora's 44.3GB across four or more says something a card count cannot, which is that sharding does not divide the requirement. Both facts are worth having and neither implies the other.

2The smaller options are documented differently too

MAGI-1 names one RTX 4090 for its 4.5B and adds at least 24GB, with 12GB for a distilled fp8 build. Open-Sora states nothing at its 256px setting, only at 768px.

So the release that speaks in cards also gives gigabytes when it matters, and the release that speaks in gigabytes leaves its cheapest configuration unmeasured.

3What a reader should take from the pair

That a multi-GPU figure without named cards is unreproducible, and a card count without a memory figure cannot be matched against hardware somebody already owns.

The register keeps each in the unit it was published in. Converting either would produce a number with no source.

  • Hardware
    H100/H800 x 8 for the 24B model, RTX 4090 x 1 for the 4.5Bstated as a count of a named cardSandAI-org, MAGI-1 repository / recorded 2026-09-22
  • Hardware
    60.3GB peak GPU memory at 768px on one GPU, and 44.3GB across four or moremeasured on H100 or H800 cardshpcaitech, Open-Sora repository / recorded 2026-09-22

4Sources

Read from github.com/SandAI-org and github.com/hpcaitech on 2026-09-22. Full entries: MAGI-1 and Open-Sora. What this column is quoting is on hardware. Nearby pairs: Two pairs, two variables, Both sides, or neither.