What the repositories say you need to run them
Eleven of the 15 families here state a GPU memory figure and four state none. The figures run from 4GB in FP16 on the smallest CogVideoX model to 80GB for the heavier Wan commands, and two repositories name a card instead of a number. As of 2026-09-22.
| Release | Stated requirement | As written |
|---|---|---|
| CogVideoX 2B | From 4GB in FP16 | From 5GB on the 5B, from 10GB on CogVideoX1.5-5B |
| FramePack 13B | 6GB minimum | For a 1-minute video at 30fps, 1800 frames |
| Allegro | 9.3G BF16 with cpu_offload | 27.5G with the offload call left out |
| MAGI-1 4.5B | At least 12GB on a distilled fp8 build | At least 24GB for the 4.5B model itself |
| SkyReels-V2 1.3B | About 14.7GB peak at 540P | Around 51.2GB for the same 540P on the 14B |
| Cosmos Predict 2.5 | 32.54 GB of GPU VRAM | Ampere, Hopper and Blackwell named as supported |
| Open-Sora 11B | 44.3GB at 768px across four or more GPUs | 60.3GB at 768px on a single H100 or H800 |
| HunyuanVideo | 45GB at 544px960px | 60GB peak at 720px1280px with 129 frames |
| Mochi 1 | About 60GB on a single GPU | At least one H100 recommended |
| Step-Video-T2V | 72.48 GB peak at 544px992px136f | 78.55 GB at 768px768px204f |
| Wan 2.2, other commands | At least 80GB of VRAM | Stated in the same repository |
| Wan 2.2 TI2V-5B | Consumer card named: a 4090 | 5-second 720P clip in under 9 minutes on a single consumer GPU |
| Stable Video Diffusion | A card named, no memory floor | About 180s on an A100 80GB card |
| LongCat-Video | No figure stated | No card named either |
| LTX-Video 13B | No minimum stated | Distilled and fp8 variants published for lower VRAM |
| Pyramid Flow | No figure stated | Offloading discussed without gigabytes |
Inclusion rule. Statements taken from the model repositories themselves. A repository that gives no figure is recorded as giving none rather than having one estimated from the parameter count. Order. By the lowest memory figure the repository states, smallest first; releases naming a card without a figure next, and releases stating neither last.
1Eleven figures, and the spread is a factor of twenty
From 4GB in FP16 on the smallest CogVideoX model to 80GB for the heavier Wan commands, by way of FramePack at 6GB, Allegro at 9.3G, a distilled MAGI-1 build at 12GB, SkyReels-V2 at about 14.7GB, Cosmos Predict 2.5 at 32.54 GB, Open-Sora at 44.3GB across four or more cards, HunyuanVideo at 45GB, Mochi 1 near 60GB and Step-Video-T2V at 72.48 GB. Every one of those is a vendor's own figure for a configuration the vendor chose to publish.
Read as a ladder the list says the obvious thing, which is that open weights have no single hardware answer. Read as a set of statements it says something less comfortable: the figures are not comparable with each other. A number for one resolution and frame count, at one precision, with offloading on or off, does not sit on the same axis as a number published for a different combination, and this register does not pretend otherwise by ranking them.
2Two numbers in one repository, and both are true
The gap between a 4090 and 80GB of VRAM in the Wan2.2 repository is the difference between a desktop and a rented server, so a reader who takes away one number has taken away the wrong one half the time. They describe different commands: the small text-to-video and image-to-video release is the one paired with the consumer card, and the heavier commands in the same repository carry the larger floor.
This register records both rather than resolving them, because resolving them would mean deciding which command a reader intends to run. What the two figures together do establish is that open weights do not have a single hardware answer, and that a claim of consumer-grade hardware attaches to a specific release rather than to a family.
3Nine minutes is a throughput figure, not a speed test
Under nine minutes for a five-second 720P clip is the timing statement in the Wan2.2 repository. Read as a ratio it is roughly a hundred seconds of compute for one second of output, which is the number that matters when the comparison is against a per-second price rather than against another card.
It is the vendor's own figure on the vendor's own chosen command, so it is the friendly case rather than an average. This site does not run the command, and a reader planning capacity should treat the figure as the shape of the answer rather than as their answer.
4Publishing smaller variants instead of a floor
The LTX-Video repository takes the opposite approach: no stated minimum, and distilled and fp8 builds published specifically to run in less memory and finish faster. That is a statement about intent rather than a specification, and it leaves the reader to find their own floor by trying.
Neither approach is better and they fail differently. A stated figure can be wrong for your command; an unstated one cannot be wrong, and cannot be planned against either.
5What the hosted rate buys instead
The alternative to hardware is a published per-second rate, and two are recorded in this register: Wan 3.0 seconds sold from about $0.06 at 480P, and LTX-2.5 seconds from $0.09. Against roughly a hundred seconds of compute per second of output, the comparison a team is actually making is a rate against a card plus the hours to keep it busy.
The versions on each side are not the same, which is the finding this whole register keeps returning to: the weights that exist are an earlier generation than the seconds that are sold. A cost comparison that ignores that is comparing a price with a different product.
6Sources
Hardware statements are quoted from each family's own repository or model card, and each one is listed below with its source and the date it was read. The hosted rates are from ltx.io and scenemixer.com, and both are recorded with their versions on the hosted pages. What counts as a published statement is on the method page.
- HardwareThe repository states TI2V-5B runs on consumer-grade graphics cards such as a 4090, generating a 5-second 720P video in under 9 minutes on a single consumer GPUvendor's own figure
- HardwareOther commands in the same repository state a requirement of at least 80GB of VRAMthe larger models
- Published modelsA 13B model, with distilled and fp8 variants published for lower VRAM and faster inferenceseveral variants
- HardwareFrom 4GB in FP16 for the 2B model, from 5GB in BF16 for the 5B, and from 10GB for CogVideoX1.5-5Bthe lowest figures the repository states
- HardwareTo generate a 1-minute video at 30fps, 1800 frames, using the 13B model, the minimal required GPU memory is 6GBa floor stated together with the output it buys
- Hardware9.3G BF16 with cpu_offload, against 27.5G if CPU offload is not enabledsingle GPU memory usage on the same run
- 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
- HardwareAbout 14.7GB peak VRAM for a 540P video on the 1.3B model, and around 51.2GB on the 14Bthe same output size on two model sizes
- HardwareThis model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architectures
- Hardware60.3GB peak GPU memory at 768px on one GPU, and 44.3GB across four or moremeasured on H100 or H800 cards
- Hardware78.55 GB peak GPU memory at 768px768px204f, 77.64 GB at 544px992px204f and 72.48 GB at 544px992px136fstated per configuration at batch size one
- HardwareSVD-XT takes about 180s on an A100 80GB carda timing on a named card with no memory floor stated
- What the hosted service sellsWan 3.0 seconds at published per-second rates, from about $0.06 at 480P
- What the hosted service sellsLTX-2.5 API seconds at $0.09 to $0.39