NodeValid

Published video-model weights, licences and hardware notes

From four gigabytes to eighty

The published figures run from 4GB in FP16 on the smallest CogVideoX model to 80GB for the heavier Wan commands. FramePack states 6GB for a 13B model and Step-Video-T2V states 72.48 GB for a 30B one. Open weights do not have one hardware answer. As of 2026-09-22.

The published memory figures, from the smallest to the largestThe figures recorded here run from four gigabytes in FP16 for the smallest model up to sixty gigabytes peak at a larger setting, with another family asking about sixty and recommending a data-centre card. A factor of twenty separates the ends of that range.Published figures, smallest firstAbout 4GBThe smallest published figure, in FP16, on the smallest model inthe register.Consumer cardsWhere the smallest releases are stated to run, on hardwarebought for other reasons.About 60GBWhere two families sit at their larger settings, one of themrecommending a data-centre card.No repository publishes throughput at the low end
Fig. 1 The ends of this range are different kinds of machine, not different budgets for one. Lower figures generally buy smaller output rather than the same output cheaper.
Published hardware figures, lowest first, with what each implies about where the model runs. Recorded 2026-09-22.
ItemWhat is publishedDetail
CogVideoX 2BFrom 4GB, FP16Laptop-class graphics card
CogVideoX 5B / 1.5-5BFrom 5GB / from 10GBDesktop card
FramePack 13B6GB, for 1800 frames at 30fpsConsumer card, RTX 30XX series and later
Allegro9.3G with cpu_offloadDesktop card, paid for in runtime
MAGI-1 4.5B12GB on a distilled fp8 build24GB for the 4.5B model itself
SkyReels-V2 1.3BAbout 14.7GB at 540PDesktop card, at the smallest variant
Cosmos Predict 2.532.54 GBAmpere, Hopper or Blackwell
Open-Sora 11B44.3GB across four or more GPUs60.3GB on a single H100 or H800
HunyuanVideo45GB at the smaller setting60GB peak at the larger one
Mochi 1About 60GB, H100 recommendedRented data-centre card
Step-Video-T2V72.48 GB at 136 frames78.55 GB at 204 frames and 768px768px
Wan 2.2, other commandsAt least 80GBRented data-centre card

Inclusion rule. Figures stated in the vendor's own repository. Community benchmarks and quantised community builds are not recorded. Order. By the lowest figure each family states.

1A factor of twenty is not a detail

Four gigabytes and eighty gigabytes are not two points on a shopping list. One is a card somebody already owns; the other is an hourly rental with a queue. Any sentence that treats open weights as a single option is hiding that spread.

The register keeps the figure attached to the specific model rather than the family, because within one family here the spread is already a factor of several.

2Lower figures usually buy smaller output

The 4GB entry point is the smallest model at the lower resolution, not the family's best output on cheap hardware. Reading down the table without reading across it produces exactly the wrong conclusion.

That is why the detail column exists and why the rows are sorted by the figure rather than by the model: the sort makes the ladder visible and the detail stops it being misread.

3What no repository publishes is throughput at the low end

Three families state a generation time: a five-second clip in under nine minutes, about 180 seconds on an A100, and a per-frame figure compared between laptops. The rest state memory and nothing else, which tells a reader whether a model will load and not whether it will finish before they lose interest.

For anyone costing local generation against a per-second rate, throughput is the missing number, and it is missing almost everywhere.

  • Hardware
    From 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 statesTHUDM, CogVideo repository / recorded 2026-09-12
  • Hardware
    Approximately 60GB VRAM on a single GPU, with at least one H100 recommendedthe vendor's own figureGenmo, Mochi repository / recorded 2026-09-12
  • Hardware
    Peak GPU memory of 60GB at 720px1280px and 129 frames, and 45GB at 544px960pxthe vendor's own figuresTencent-Hunyuan, HunyuanVideo repository / recorded 2026-09-12
  • Hardware
    To 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 buyslllyasviel, FramePack repository / recorded 2026-09-22
  • Hardware
    9.3G BF16 with cpu_offload, against 27.5G if CPU offload is not enabledsingle GPU memory usage on the same runrhymes-ai/Allegro on Hugging Face / recorded 2026-09-22
  • Hardware
    For 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 buildSandAI-org, MAGI-1 repository / recorded 2026-09-22
  • Hardware
    About 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 sizesSkyworkAI, SkyReels-V2 repository / recorded 2026-09-22
  • Hardware
    This model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architecturesnvidia/Cosmos-Predict2.5-2B on Hugging Face / 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
  • Hardware
    78.55 GB peak GPU memory at 768px768px204f, 77.64 GB at 544px992px204f and 72.48 GB at 544px992px136fstated per configuration at batch size onestepfun-ai, Step-Video-T2V repository / recorded 2026-09-22
  • Hardware
    Other commands in the same repository state a requirement of at least 80GB of VRAMthe larger modelsWan-Video, Wan2.2 repository / recorded 2026-09-12

4Open weights, hosted

Instead of local hardware for “The VRAM ladder”, SceneMixer runs hosted models; its guides describe the result:

SceneMixer turns a novel into a watchable short-drama series in one pipeline: upload the novel (Chinese or English) → AI extracts characters, scenes, and an episode-by-episode structure → generates consistent character and scene art → writes storyboard scripts → renders video segment by segment (16:9 by default, 9:16 in one setting) → composites full episodes.

SceneMixer, Turn a Novel into a Video Series with AI (read 2026-10-13)

The Seedance 2.0 tiers render segments up to 15 seconds; use a Seedance 2.5 or Wan tier for 30-second takes.

SceneMixer, Seedance 2.0 Fast — “Does Fast render 30-second takes?” (read 2026-10-13)

5Sources

Repository statements and hosted rates are linked from the model pages and the method page, checked 2026-09-22. Hardware figures are set out on requirements. Nearby: Naming the card, How long a clip, Nobody sells these.