Wan 2.2 5B, a 5-second 720p video
- RTX PRO 6000 96 GB, $2.42/h
- 74 s · $0.05
- RTX 6000 Ada 48 GB, $1.37/h
- 200 s · $0.076
2.7 times faster, and 35% less for one video.
Wan 2.2
Wan 2.2 (Text to video, A14B) needs 14.5 GB of VRAM in ComfyUI, at the diffusion stage of its own template, by file sizes read 3 November 2026. The smallest card that loads it with room to work is the RTX 4090 (24 GB), and the template downloads 35.6 GB.
- Made by
- Wan (Alibaba)
- Licence
- Apache 2.0
- Released
- July 2025
- Sizes read
- 3 November 2026
Needs
14.5 GB
At the diffusion stage, with the FP8 scaled file ComfyUI's template loads. The template downloads 35.6 GB.
Computed 3 November 2026
Smallest card
24 GB
Computed 3 November 2026
Cheapest clip we know
$0.30
RunPod's API: a 5-second 1280×720 clip, on its model and settings, not your workflow.
Read 3 November 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Needs
14.5 GB
At the diffusion stage, with the FP8 scaled file ComfyUI's template loads. The template downloads 35.6 GB.
Computed 3 November 2026
Smallest card
24 GB
Computed 3 November 2026
Runs on Comfy Cloud?
Yes
ComfyUI's template for it uses built-in nodes only, and its largest stage, 14.5 GB, fits the 96 GB RTX PRO 6000.
Read 3 November 2026
Cheapest clip we know
$0.05
On Comfy Cloud: 11 credits for a 5-second 640×640 clip, by Comfy's own figure.
Read 3 November 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Needs
11.4 GB
At the diffusion stage, with the FP16 file ComfyUI's template loads. The template downloads 18.2 GB.
Computed 3 November 2026
Cheapest clip we know
$0.05
Our run on a rented RTX PRO 6000: Wan 2.2 5B, a 5-second 720p video, in 74 s.
Measured 20 October 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Needs
16.9 GB
At the diffusion stage, with the INT8 ConvRot file ComfyUI's template loads. The template downloads 24.9 GB.
Computed 3 November 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Which file on your card
For each memory size, the best file for that card: the highest-precision diffusion file that loads with room to work (one that only just fits when nothing else does), and what is left for working memory while it samples. On a larger card that can be above the file ComfyUI's template loads, which the answer above is about. "Streams" means even the smallest file is larger than the card: ComfyUI still runs it from system RAM, more slowly. FP4 files go only to the Blackwell cards that run them natively.
Text to video, A14B · two experts, one on the card at a time
ComfyUI's template for Text to video, A14B: FP8 scaled diffusion model, FP8 scaled text encoder. Stages: text encoder 6.7 GB, diffusion 14.5 GB. Download 35.6 GB.
Computed 3 November 2026Image to video, A14B · two experts, one on the card at a time
ComfyUI's template for Image to video, A14B: FP8 scaled diffusion model, FP8 scaled text encoder. Stages: text encoder 6.7 GB, diffusion 14.5 GB. Download 35.6 GB.
Computed 3 November 2026Text and image to video, 5B
ComfyUI's template for Text and image to video, 5B: FP16 diffusion model, FP8 scaled text encoder. Stages: text encoder 6.7 GB, diffusion 11.4 GB. Download 18.2 GB.
Computed 3 November 2026Animate 2
ComfyUI's template for Animate 2: INT8 ConvRot diffusion model, FP8 scaled text encoder. Stages: text encoder 8.0 GB, diffusion 16.9 GB. Download 24.9 GB.
Computed 3 November 2026Check another card, or pick a file by hand, in the ComfyUI VRAM checker. Built your own workflow around it, with LoRAs or custom nodes? Drop it into the ComfyUI workflow checker.
Where to run it
Text to video, A14B
Your own GPU
24 GB or more
Loads ComfyUI's template with room to work. 16 GB loads it, tight.
Computed 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
A per-clip API
$0.30 a clip
RunPod sells Wan 2.2 by the clip, 5 seconds at 1280×720: its model and settings, not your workflow.
Read 3 November 2026
On-demand prices before tax, read 19 October to 3 November 2026.
Image to video, A14B
Your own GPU
24 GB or more
Loads ComfyUI's template with room to work. 16 GB loads it, tight.
Computed 3 November 2026
Comfy Cloud
$4.54 an hour of generation
A 96 GB RTX PRO 6000, billed only while a workflow runs. Plans from $20 a month.
Read 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
A per-clip API
$0.30 a clip
RunPod sells Wan 2.2 by the clip, 5 seconds at 1280×720: its model and settings, not your workflow.
Read 3 November 2026
On-demand prices before tax, read 19 October to 3 November 2026.
Text and image to video, 5B
Your own GPU
16 GB or more
Loads ComfyUI's template with room to work.
Computed 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
A per-clip API
$0.30 a clip
RunPod sells Wan 2.2 by the clip, 5 seconds at 1280×720: its model and settings, not your workflow.
Read 3 November 2026
On-demand prices before tax, read 19 October to 3 November 2026.
Animate 2
Your own GPU
24 GB or more
Loads ComfyUI's template with room to work.
Computed 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
On-demand prices before tax, read 19 October 2026.
What we measured
Wan 2.2 5B, a 5-second 720p video, on rented cards in Finland on 20 October 2026. With every model loaded, the 121-frame video took about 31 GB, against a largest stage of 11.4 GB: working memory comes on top.
Measured 20 October 2026 How we measure
About the model
Wan 2.2 is Wan's open video model. Alibaba's Wan 2.2, from July 2025, remains the newest Wan with open weights. The A14B models are two 14-billion-parameter experts that take turns, one for the rough layout and one for detail. TI2V-5B is a single smaller model with a more compressed VAE, and Animate 2 (August 2026) is the character-animation model in the same family.
Text to video, A14B: 14B · Image to video, A14B: 14B · Text and image to video, 5B: 5B · Animate 2: 14B Text encoder: UMT5-XXL Apache 2.0 Released July 2025
Good at
- Apache 2.0, so commercial use is open.
- 480p and 720p clips of about five seconds.
- LoRA training on 24 GB cards, documented by musubi-tuner and ai-toolkit.
Watch out for
- Two experts mean two downloads of the same size, even though only one is on the card at a time.
- Wan's own script for the A14B models asks for an 80 GB GPU; ComfyUI gets by with far less by moving models in and out.
- Wan 2.5, 2.6, 2.7 and 3.0 are hosted only, with no open weights.
Every Wan 2.2 file ComfyUI loads
Sizes are the exact byte counts Hugging Face reports, read 3 November 2026. Files marked "template" are the ones ComfyUI's official workflow downloads.
| Part | Precision | Size | Version | File |
|---|---|---|---|---|
| Diffusion model | FP16 | 28.58 GB | Text to video, A14B | wan2.2_t2v_high_noise_14B_fp16.safetensors |
| Diffusion model | FP8 scaled · template | 14.29 GB | Text to video, A14B | wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors |
| Diffusion model | FP16 | 28.58 GB | Image to video, A14B | wan2.2_i2v_high_noise_14B_fp16.safetensors |
| Diffusion model | FP8 scaled · template | 14.29 GB | Image to video, A14B | wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors |
| Diffusion model | FP16 · template | 10.00 GB | Text and image to video, 5B | wan2.2_ti2v_5B_fp16.safetensors |
| Diffusion model | BF16 | 32.79 GB | Animate 2 | wan_animate_2_bf16.safetensors |
| Diffusion model | INT8 ConvRot · template | 16.65 GB | Animate 2 | wan_animate_2_int8_convrot.safetensors |
| Text encoder | FP16 | 11.37 GB | all | umt5_xxl_fp16.safetensors |
| Text encoder | FP8 scaled · template | 6.74 GB | all | umt5_xxl_fp8_e4m3fn_scaled.safetensors |
| VAE | BF16 · template | 0.25 GB | Text to video, A14B | wan_2.1_vae.safetensors |
| VAE | BF16 · template | 0.25 GB | Image to video, A14B | wan_2.1_vae.safetensors |
| VAE | FP16 · template | 1.41 GB | Text and image to video, 5B | wan2.2_vae.safetensors |
| VAE | BF16 · template | 0.25 GB | Animate 2 | Wan2_1_VAE_bf16.safetensors |
| Vision encoder | FP16 · template | 1.26 GB | Animate 2 | clip_vision_h.safetensors |
What Wan says it needs
- "This command can run on a GPU with at least 80GB VRAM." (A14B models, Wan's own script, source)
- "This command can run on a GPU with at least 80GB VRAM." (A14B models, Wan's own script, source)
- "This command can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)." (TI2V-5B at 720p, Wan's own script with offloading, source)
Those figures and this page measure different things. A vendor's figure covers its own script at its default resolution and length, working memory included. This page counts the weights each stage holds, which is exact, and leaves the working memory as the space that remains, because it depends on your resolution, frame count and attention kernel and nobody publishes it per model. If generation runs out of memory with a file that loads, lower the resolution or the frame count before stepping down a precision. ComfyUI's README says it "can run even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights (source), so "streams" means slower, not impossible.
Wan 2.5, 2.6, 2.7 and 3.0 have no open weights (checked 2026-10-02). None has open weights on Hugging Face (Wan-AI), ModelScope or GitHub (Wan-Video). Wan offers them through Alibaba Cloud Model Studio, the Wan website, the Qwen app and Qwen Cloud, and has not said they will be released.
LoRA training for Wan 2.2
Only configurations a trainer's own documentation gives a VRAM figure for. Settings change the figure a great deal, so read the setting column before trusting the number.
Questions
- How much VRAM does Wan 2.2 need?
- In ComfyUI's own workflow for Text to video, A14B, the largest stage is the diffusion stage at 14.5 GB, and the whole workflow downloads 35.6 GB. Working memory for generation comes on top and grows with resolution and length. Wan itself says: "This command can run on a GPU with at least 80GB VRAM." (A14B models, Wan's own script).
- Can I run Wan 2.2 on 8 GB of VRAM?
- Not fully. The smallest file here, FP8 scaled at 14.3 GB, needs 14.5 GB at its largest stage, more than 8 GB holds. ComfyUI will still run it by streaming weights from system RAM, which works but is slower.
- Can I run Wan 2.2 on 12 GB of VRAM?
- Not fully. The smallest file here, FP8 scaled at 14.3 GB, needs 14.5 GB at its largest stage, more than 12 GB holds. ComfyUI will still run it by streaming weights from system RAM, which works but is slower.
- Can I run Wan 2.2 on 16 GB of VRAM?
- Yes, with the FP8 scaled file: its largest stage, the diffusion stage, is 14.5 GB, which loads on 16 GB with 1.5 GB left for working memory, a tight fit. That is little room to generate at full resolution.
- Which Wan 2.2 file should I download?
- The highest precision that loads fully on your card. On a 24 GB card that is the FP8 scaled file (14.3 GB). ComfyUI's template downloads FP8 scaled with the FP8 scaled text encoder. FP4 files (NVFP4) are made for Blackwell cards (RTX 50, RTX PRO 6000, DGX Spark); on older cards, use FP8 or INT8.
- Can I run Wan 2.5, 2.6, 2.7 or 3.0 locally?
- No. None of them has open weights: they are not on Hugging Face, ModelScope or Wan's GitHub, and Wan offers them only through Alibaba Cloud Model Studio, the Wan website, the Qwen app and Qwen Cloud. Wan 2.2, including Animate 2, is the newest you can download.
- Files and sizes: Comfy-Org/Wan_2.2_ComfyUI_Repackaged on Hugging Face, read from the Hugging Face API. Template files: ComfyUI's Wan 2.2 guide.
- Model and licence: Wan-AI/Wan2.2-T2V-A14B and the Apache 2.0.
- Card memory: manufacturers' specifications, as on each GPU page. The checker covers NVIDIA cards; ComfyUI also runs on AMD and Apple silicon, where file formats and speed differ and we have not sized them.
- Fit rule: a stage loads fully under 85% of usable memory and is tight up to 95%, the same margins the LLM pages use.
- Prices: each provider's own pricing page, on the dates beside them. Comfy Cloud's answer comes from ComfyUI's own template for that variant, which uses built-in nodes only, and shows only for the variants whose template we checked.