Home → Video generation
VRAM for AI video generation
Video models are the heaviest of the creative workloads. The table below uses official README requirements where they exist, and labels the rest.
Model requirements
| Model | VRAM | Note | Source |
|---|---|---|---|
| Wan 2.1 T2V-1.3B | ≈ 8.19GB | The only official tier that natively fits a 24GB consumer card at low tiers | Official README (github.com/Wan-Video/Wan2.1) |
| Wan 2.2 TI2V-5B | 24GB @ 720p | Runs natively on an RTX 4090 | Official README (github.com/Wan-Video/Wan2.2) |
| Wan 2.1 14B | ≈ 40GB+ (quantized: 24GB workable) | Precise quantized figure not verified | Official range + community reports |
| HunyuanVideo | 45GB (540p) / 60GB (720p) | Multi-GPU or rented datacenter-class territory | Official README (github.com/Tencent/HunyuanVideo) |
| LTX-Video | ≈ 8–12GB | Community-reported figures | Community (labeled as such) |
Retrieved October 2026. "Community" rows are directionally useful but not official.
The 24GB question
A 24GB consumer card (RTX 3090 / 4090) runs Wan 2.1 1.3B comfortably and Wan 2.2 TI2V-5B at 720p natively. It does not run the 14B class at full precision or HunyuanVideo — those need 40GB+, which in practice means quantized builds on a 24GB card or a rented larger card.
Render time context
- LTX-Video generates in near-real-time (seconds-class for short clips).
- Wan 2.2 5B takes on the order of ~4 minutes per 5-second clip (official figure).
- 14B quantized runs roughly 15–60 minutes; HunyuanVideo 30 minutes and up.
Directional figures from official docs and community reports, retrieved Oct 2026. No dollar-cost conversions are implied on this page.