Which GPU should you use for Wan video generation?

Short answer: Choose the card by Wan model and by tool, because the official script and ComfyUI need very different amounts of memory.

  • TI2V-5B: a 24 GB card such as the RTX 4090 is the class the Wan2.2 README names: “This command can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)” with its offload flags. In ComfyUI the 5B model “should fit well on 8GB vram with the ComfyUI native offloading”.
  • 14B models (T2V-A14B, I2V-A14B, S2V-14B): the README’s single-GPU commands say “This command can run on a GPU with at least 80GB VRAM”. ComfyUI’s FP8 files and offloading bring that down, but the ComfyUI docs give no 14B figure.

Sources: Wan2.2 README, ComfyUI Wan 2.2 tutorial, both read 2026-11-04.

Which Wan models are open?

Version Open weights License
Wan 2.1 (T2V-1.3B, T2V-14B, I2V-14B, FLF2V-14B, VACE) Yes Apache 2.0
Wan 2.2 TI2V-5B, T2V-A14B, I2V-A14B (“Jul 28, 2025”) Yes Apache 2.0
Wan 2.2 S2V-14B (audio-driven) Yes Apache 2.0
Wan 2.2 Animate-14B, Wan-Animate-2 (2026-08-07) Yes Apache 2.0
Wan 2.5, Wan 3.0 No weights in the Wan-AI Hugging Face organisation Hosted access only

The model list was checked against the Wan-AI organisation on Hugging Face and the Wan2.1 and Wan2.2 READMEs, read 2026-11-04. If you need Wan 3.0, there is nothing to put on a GPU: use CloudGPU’s model API instead. The video models there include wan-3.0 and wan-3.0-prime, which you call with POST /v1/videos (API docs).

Which card for which Wan model

The README and ComfyUI wording below was read 2026-11-04. The Deploy page shows which cards you can start now.

Model What the official sources say Card class to start with
Wan 2.1 T2V-1.3B “requires only 8.19 GB VRAM”; “a 5-second 480P video on an RTX 4090 in about 4 minutes (without optimization techniques like quantization)” 24 GB: RTX 3090, RTX 4090, RTX 4090D
Wan 2.2 TI2V-5B (ComfyUI) “should fit well on 8GB vram with the ComfyUI native offloading” 24 GB: RTX 4090, RTX 4090D; 32 GB if you want 720p without offloading: RTX 5090, RTX 5090D, RTX 4080 SUPER 32G
Wan 2.2 TI2V-5B (official generate.py) “at least 24GB VRAM (e.g, RTX 4090 GPU)” with --offload_model True --convert_model_dtype --t5_cpu; “Without specific optimization, TI2V-5B can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU” 24 GB or more
Wan 2.2 T2V-A14B / I2V-A14B (ComfyUI FP8) Two experts of 14.29 GB each plus the text encoder; the ComfyUI docs give no VRAM figure for 14B 32–48 GB: RTX 5090, RTX 4090 48G, RTX 4090D 48G, RTX A6000
Wan 2.2 T2V-A14B / I2V-A14B (official script) “This command can run on a GPU with at least 80GB VRAM.” 80 GB or more: A100 80G, A800 80G, RTX 6000D, RTX PRO 6000 96G
Wan 2.2 S2V-14B “This command can run on a GPU with at least 80GB VRAM.” 80 GB or more
Wan-Animate-2 14B “The default settings in this repository are tuned for 8× A800 GPUs and support 720P video generation. We have also tested 480P generation on 2× A800 GPUs.” Several 80 GB cards for the reference setup

The Wan2.2 README also says: with “at least 80GB VRAM, you can remove the --offload_model True, --convert_model_dtype and --t5_cpu options to speed up execution”. A larger card lets you drop the offload flags, which are what make the smaller cards fit.

What changes the memory you need

  • Resolution and frame count. In ComfyUI, length and the latent size change memory the most. The ComfyUI docs set “a relatively small size … to prevent low VRAM users from consuming too many resources” and say “If you have enough VRAM, you can try a resolution around 720P”.
  • Precision. FP8 14B experts are 14.29 GB each, against 28.58 GB each in FP16 (Hugging Face file sizes, Comfy-Org Wan 2.2 repack, read 2026-11-04).
  • Offloading to host RAM. Offloading trades VRAM for system memory, so check the RAM per GPU on /deploy.

If you run out of memory, the README says “you can use the --offload_model True, --convert_model_dtype and --t5_cpu options to reduce GPU memory usage.”

Training a Wan LoRA

For consistent characters or a house style, train a LoRA. The musubi-tuner docs say: “VRAM: 12GB or more recommended for image training, 24GB or more for video training” and “Main Memory: 64GB or more recommended” (musubi-tuner, read 2026-11-04). Step-by-step instructions are in how to train a Wan 2.2 LoRA with musubi-tuner.

Cost

At 2026-11-05 07:55 (UTC+8): RTX 4090 $0.39/hour, RTX 5090 $0.69/hour, RTX 4090 48G $0.69/hour, A100 80G $1.59/hour. A 30-minute test session costs $0.20 on the RTX 4090 and $0.34 on the RTX 5090. Billing is per minute, rounded up, with a 1-minute minimum, and it stops when you destroy the instance. Current rates are on /pricing.

To set it up, follow how to run Wan 2.2 video generation on a cloud GPU. Related: Which GPU for MiniMax H3 · AI short film production: which GPU for each step · Cloud GPU for AI film and video production

Frequently asked questions

What GPU do I need for Wan 2.2 TI2V-5B?

The Wan2.2 README says its own script runs TI2V-5B “on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)” with offload flags, and the ComfyUI tutorial says the 5B model “should fit well on 8GB vram with the ComfyUI native offloading” (read 2026-11-04).

What GPU do I need for Wan 2.2 14B?

The Wan2.2 README’s single-GPU commands for T2V-A14B, I2V-A14B and S2V-14B say “This command can run on a GPU with at least 80GB VRAM”. ComfyUI’s FP8 files, 14.29 GB per expert, and its offloading lower that, but the ComfyUI docs give no 14B figure (read 2026-11-04).

Is Wan 3.0 open source?

No Wan 2.5 or Wan 3.0 weights are published in the Wan-AI Hugging Face organisation, which lists Wan 2.1, Wan 2.2, Wan-Dancer and Wan-Animate-2 (read 2026-11-04). CloudGPU’s model API offers wan-3.0 and wan-3.0-prime through POST /v1/videos.

Can I use Wan output commercially?

The Wan 2.1 and Wan 2.2 models are released under Apache 2.0, according to their READMEs and Hugging Face pages (read 2026-11-04). Check the licence on the model page for the files you use.

Does more VRAM make Wan faster?

It lets you drop the offload options. The Wan2.2 README says that on a GPU with at least 80GB VRAM “you can remove the --offload_model True, --convert_model_dtype and --t5_cpu options to speed up execution” (read 2026-11-04).

Sources (read 2026-11-04, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing.

Last updated: 2026-11-04 (Asia/Shanghai).