MAGI-1: eight H100s or one 4090, by size
The repository states H100 or H800 x 8 for the 24B model and RTX 4090 x 1 for the 4.5B. It adds that the 4.5B needs at least 24GB, and that a distilled fp8 configuration works with at least 12GB. As of 2026-09-22.
| Column | What the pages state |
|---|---|
| Hardware, in the vendor's own words | H100 or H800 x 8 for the 24B |
| The detail published beside it | One RTX 4090 for the 4.5B; 12GB for a distilled fp8 build |
| Resolution, from the same pages | 720 x 720 as the 4.5B default |
| Length, from the same pages | No ceiling recorded |
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.
1Counting cards says more than counting gigabytes
Eight H100s is not just 640GB of memory. It is a node with a fast interconnect, a driver stack and formats the run was exercised against. Written as a memory total it would be unreproducible.
One RTX 4090 works the same way in the other direction: consumer memory, consumer bandwidth, and a format set that is not the same as a data-centre card's.
2Three floors for two sets of weights
At least 24GB for the 4.5B model itself, at least 12GB for a distilled fp8 build of it, and eight data-centre cards for the 24B. The middle figure is about software, not weights.
That is the clearest statement in the register that a memory requirement describes a build. The parameter count does not change between 24GB and 12GB; the numeric format does.
3What the resolution default tells you
The 4.5B default is 720x720, described as set to accelerate inference, with arbitrary resolutions supported. So the hardware figures belong to a square default rather than to the output a reader may want.
Length is unbounded in the repository: frames are an argument with no stated maximum. Neither of the qualifying columns pins these figures down as tightly as the card counts suggest.
- HardwareFor 4.5B models, any machine with at least 24GB of GPU memory is sufficient, and a distilled fp8 configuration works on GPUs with at least 12GBa floor for the model and a lower one for a build
- HardwareH100/H800 x 8 for the 24B model, RTX 4090 x 1 for the 4.5Bstated as a count of a named card
4Sources
Every figure above is the vendor's own, taken from github.com/SandAI-org on 2026-09-22 with nothing measured here. The rest of the release is on its family entry, the whole column is on hardware, and all memory figures together are on requirements. Nearby: 32.54 GB, three architectures, No card, no figure.