Home › Image & video models › Wan 2.1
Wan 2.1 VRAM requirements (1.3B & 14B)
official README bullet: T2V-1.3B requires only 8.19 GB VRAM (480P; matches the 1×4090 measured cell 8.19 GB). Below is the complete official "Computational Efficiency" table — every cell is one official 'total time (s) / peak GPU memory (GB)' pair, 720P × 81 frames (480P for 1.3B), card counts 1/2/4/8:
| GPU | task | res | 1 GPU | 2 GPU | 4 GPU | 8 GPU |
|---|---|---|---|---|---|---|
| RTX 4090 | T2V-1.3B | 480P | 261.4 / 8.19 GB | 188.5 / 18.1 GB | 120.8 / 14.1 GB | 112.3 / 12.2 GB |
| H20 | T2V-14B | 720P | 6935.5 / 69.1 GB | 3488.4 / 53.9 GB | 1775.8 / 37.9 GB | 980.5 / 29.9 GB |
| H20 | I2V-14B | 720P | 5494.8 / 76.7 GB | 2756.6 / 59.8 GB | 1417.0 / 41.9 GB | 778.2 / 32.9 GB |
| A800/A100 | T2V-14B | 720P | 3342.6 / 69.1 GB | 1722.0 / 52.5 GB | 886.1 / 36.6 GB | 469.9 / 28.6 GB |
| A800/A100 | I2V-14B | 720P | 2715.5 / 76.7 GB | 1371.7 / 58.5 GB | 716.1 / 40.6 GB | 389.7 / 31.6 GB |
| H800/H100 | T2V-14B | 720P | 1837.9 / 69.1 GB | 945.1 / 53.9 GB | 497.0 / 37.9 GB | 287.9 / 29.9 GB |
| H800/H100 | I2V-14B | 720P | 1491.6 / 76.7 GB | 753.9 / 59.8 GB | 403.9 / 41.9 GB | 238.8 / 32.9 GB |
Table cells are official 'total time (s) / peak GPU memory (GB)'. 14B 1-GPU tests used --offload_model True; prompt extension (local Qwen) adds its own VRAM on top. No official 14B-on-consumer-card row exists in the table — we do not guess it. Source: Wan-Video/Wan2.1 official README + efficiency figure assets/comp_effic.png (retrieved 2026-10-20); full matrix transcribed verbatim into our repo archive data/wan21_efficiency_official.txt.
What the matrix actually says (nobody tabulates this)
- 1.3B on consumer cards: one 4090 peaks at 8.19 GB — fine on 16-24 GB. But look across the row: adding GPUs raises peak memory (18.1 GB at 2 cards) before it falls (12.2 GB at 8). Multi-GPU for 1.3B buys speed, never headroom.
- 14B wants 80 GB class: single-card peak 69.1 GB (T2V) / 76.7 GB (I2V). The 8-GPU column only reaches 28.6-31.6 GB for most rows (32.9 GB worst case) — real-world that's 32-48 GB class with little slack, not a 24 GB escape hatch.
- No official 4090/14B cell exists — quantized community builds fill that gap commercially, but we publish zero guesses about their peaks.
Cost of ONE 720P video at any card count (the part rent-billing hides)
Arithmetic on official timings × our cited same-class per-card rates — assumption: N cards at the same per-card rate (multi-GPU instance pricing not independently verified; modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending):
| class (T2V-14B) | rate | 1×card | 2×cards | 4×cards | 8×cards |
|---|---|---|---|---|---|
| A800/A100 class | RunPod A100 80GB @ $1.59/h | $1.48 | $1.52 | $1.57 | $1.66 |
| H800/H100 class | RunPod H100 80GB @ $3.49/h | $1.78 | $1.83 | $1.93 | $2.23 |
Read this as a rental buyer: the cost per video is nearly flat ($1.48–$2.23 across every configuration) while wall-clock time falls about 7× within a card class (more only if you also move to a faster class). Cloud billing sells GPU-hours, so multi-GPU is a speed purchase, not a savings. Cheapest cited 80 GB class rental: RunPod A100 80GB @ $1.59/h. Buying a 24 GB card only wins if you render enough volume to amortise it — that exact break-even is on our payback calculator.
Measured a 14B peak on a consumer card with a quant build? Post it — we cite user measurements with handles. LLM side of the family: model VRAM pages · cost calculator · sibling: FLUX VRAM (no official number — honest math).
FAQ
How much VRAM does Wan 2.1 14B need?
Per the official efficiency table: 720P text-to-video peaks at 69.1 GB on one GPU (image-to-video 76.7 GB) with --offload_model True. Spreading over 8 GPUs lowers the per-card peak only to 28.6-32.9 GB depending on task and card class — still no 24 GB row. Source: Wan-Video/Wan2.1 official README + efficiency figure assets/comp_effic.png (retrieved 2026-10-20); full matrix transcribed verbatim into our repo archive data/wan21_efficiency_official.txt.
Can Wan 2.1 run on an RTX 4090 (24 GB)?
Officially: yes for T2V-1.3B — the README claims 8.19 GB and their own 4090 test measured exactly that. For 14B there is no official consumer-card row; community GGUF/fp8 quant builds exist but carry no official figure, so we publish none — run it and tell us your peak.
How much does one Wan 2.1 14B video cost to rent?
Class-matched math, not a mashup: official A800/A100 timing 3342.6 s × our cited A100-80GB rate ($1.59/h, RunPod) ≈ $1.48; official H800/H100 timing 1837.9 s × cited H100 rate ($3.49/h, RunPod) ≈ $1.78. Electricity-free either way. Rates: modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending.
Does renting 8 GPUs make each video cheaper?
No — the official timings make total GPU-hours roughly constant. Same per-card rate × card count: A100-class $1.48 (1 GPU) → $1.66 (8 GPUs); H100-class $1.78 → $2.23. You pay a little MORE for 8 cards and buy 7× less waiting. Assumes N cards at the same per-card rate (multi-GPU instance prices not independently verified).
What eats the extra memory beyond weights?
Umt5-XXL text encoder (can be pinned to CPU with --t5_cpu), the Wan-VAE decode pass, and activations at 720P×81 frames. Prompt extension with a local Qwen model adds its own footprint on top.
All timing/memory numbers on this page are official Wan2.1 figures retrieved 2026-10-20 (full transcription archived with this site's source); rental rates labelled with provider & check date. Put the headline number in your README: Wan 2.1 badge →