MAGI-1: what the repository publishes
MAGI-1 states hardware as machine shapes rather than memory: H100 or H800 times eight for the 24B model, one RTX 4090 for the 4.5B. It adds that the 4.5B needs at least 24GB, and that a distilled fp8 build runs on 12GB. The licence is Apache 2.0. As of 2026-09-22.
| Field | What is stated | Detail |
|---|---|---|
| Licence | Apache License 2.0 | Stated in the repository |
| Size | 24B and 4.5B models | Both published, each with its own hardware note |
| Hardware | H100 or H800 x 8 for the 24B | One RTX 4090 for the 4.5B; 12GB for a distilled fp8 build |
| Resolution | 720 x 720 as the 4.5B default | Set for speed; arbitrary resolutions described as supported |
| Length | No ceiling recorded | Frames are an argument with no stated maximum |
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.
1Eight cards of one kind, or one card of another
The hardware column of the model table reads H100/H800 x 8 and RTX 4090 x 1. That is a count of a named card, which is the one hardware unit in this register that converts into a rented price without an intervening decision by the reader.
It also describes the shape of the problem rather than a threshold. A 24B model wanting eight data-centre cards is not a model that one larger card unlocks, and no memory figure would have carried that as directly as a multiplier does.
2Two floors for the smaller model, and the lower one belongs to a build
Any machine with at least 24GB of GPU memory is described as sufficient for the 4.5B models, and a distilled fp8 configuration with a window size of one is said to work on GPUs with at least 12GB. The second figure is a property of that specific build, not of the model.
Halving a memory floor by distilling and quantising is a routine move, and this repository is unusual in publishing the result of it as a number. Elsewhere in this register the same move shows up only as the existence of a variant, with the hardware left for a reader to discover by trying.
3A resolution stated as a default chosen for speed
The default resolution for the 4.5B model is 720 by 720, and the repository says it was set that way to accelerate inference. Arbitrary resolutions are described as supported, so the figure is a starting point rather than a capability limit.
No clip length is stated anywhere. Frames are passed as an argument and no maximum is attached to them, so the length cell is empty on the entry whose hardware cell is the fullest in this register.
4The same field, across every family
| Family | Licence |
|---|---|
| Allegro | apache-2.0 |
| CogVideoX | Split by model size |
| Cosmos Predict 2.5 | NVIDIA Open Model License |
| FramePack | Apache 2.0 |
| HunyuanVideo | tencent-hunyuan-community |
| LongCat-Video | MIT License |
| LTX-Video | Apache 2.0 |
| MAGI-1 | Apache License 2.0 |
| Mochi 1 | Apache 2.0 |
| Open-Sora | Apache 2.0 |
| Pyramid Flow | stabilityai-ai-community |
| SkyReels-V2 | skywork-license |
| Stable Video Diffusion | stable-video-diffusion-community |
| Step-Video-T2V | MIT |
| Wan | Apache 2.0 |
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.
| Family | Size |
|---|---|
| Allegro | VAE 175M and DiT 2.8B |
| CogVideoX | 2B and 5B, carried in the model names |
| Cosmos Predict 2.5 | 2,059,174,912 parameters |
| FramePack | 13B models |
| HunyuanVideo | Over 13 billion parameters |
| LongCat-Video | 13.6B parameters |
| LTX-Video | A 13B model |
| MAGI-1 | 24B and 4.5B models |
| Mochi 1 | 10 billion parameters |
| Open-Sora | An 11B model |
| Pyramid Flow | No parameter count recorded |
| SkyReels-V2 | 1.3B, 5B and 14B variants |
| Stable Video Diffusion | 2B params |
| Step-Video-T2V | 30 billion parameters |
| Wan | A 5B text-to-video and image-to-video model |
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.
| Family | Hardware |
|---|---|
| Allegro | 9.3G BF16 with cpu_offload |
| CogVideoX | From 4GB in FP16 on the 2B |
| Cosmos Predict 2.5 | 32.54 GB of GPU VRAM |
| FramePack | 6GB minimum for 1800 frames at 30fps |
| HunyuanVideo | 60GB peak at 720px1280px, 129 frames |
| LongCat-Video | No memory figure recorded |
| LTX-Video | No minimum stated |
| MAGI-1 | H100 or H800 x 8 for the 24B |
| Mochi 1 | About 60GB VRAM on a single GPU |
| Open-Sora | 60.3GB peak at 768px on one GPU |
| Pyramid Flow | No memory figure recorded |
| SkyReels-V2 | About 14.7GB peak on the 1.3B at 540P |
| Stable Video Diffusion | About 180s on an A100 80GB card |
| Step-Video-T2V | 78.55 GB peak at 768px768px204f |
| Wan | A consumer 4090 on the 5B release |
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.
- LicenceApache License 2.0stated in the repository
- Model sizeMAGI-1-24B and MAGI-1-4.5Btwo published families of weights
- HardwareH100/H800 x 8 for the 24B model, RTX 4090 x 1 for the 4.5Bstated as a count of a named card
- 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
- ResolutionThe default resolution for the 4.5B model is set to 720x720a default chosen to accelerate inference
5Sources
Read from github.com/SandAI-org 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: SkyReels-V2. Next family: Cosmos Predict 2.5.