HunyuanVideo VRAM: 45 GB at 544p, 60 GB at 720p, 80 GB advised
The HunyuanVideo README lists 45 GB for 544x960 and 60 GB for 720x1280 at 129 frames, tested on one 80 GB GPU. FP8 saves about 10 GB. Table and hosted route.

The original HunyuanVideo README states its GPU peak memory plainly: 60 GB for 720x1280 at 129 frames and 45 GB for 544x960 at 129 frames. It says the model was tested on a single 80 GB GPU, names 80 GB as the recommendation, and requires an NVIDIA GPU with CUDA on Linux. The FP8 weights save about 10 GB. No 24 GB consumer card meets those numbers without extra tooling. Numbers are from the README read on 2026-10-08.
The README table
Memory is for one generation of 129 frames. A longer clip is not covered by the table.
| Setting (height x width x frames) | GPU peak memory | Minimum per README | With FP8 weights (arithmetic: minus about 10 GB) |
|---|---|---|---|
| 720px x 1280px x 129f | 60 GB | 60 GB | about 50 GB |
| 544px x 960px x 129f | 45 GB | 45 GB | about 35 GB |
What the README requires
Three lines in the system requirements shape the buying decision.
- An NVIDIA GPU with CUDA support is required.
- The model is tested on a single 80 GB GPU; 80 GB is recommended for better quality.
- The tested operating system is Linux, and the install guide is for Linux.
What it means for a normal workstation
A 24 GB card is below even the 544p row, so people use third-party low-memory ports, which the README links under community projects. Those are not the vendor's measured figures, so budget testing time. The FP8 file saves about 10 GB by the README's own statement; the 'about 35 GB' and 'about 50 GB' cells above are subtraction, not measurements.
Rented cards with 80 GB are the path the README itself assumes. If you only need occasional clips, the question becomes how many clips per month pay for an 80 GB instance's idle time.
Reading the numbers against your budget
The table is peak memory, so a card with exactly 60 GB leaves no room for a desktop session or a second process. The README's recommendation of 80 GB is the practical floor for 720p. If you rent by the hour, a node with an 80 GB card is the unit; two 40 GB cards do not pool into 80 GB unless you use the README's multi-GPU parallel path, which is a separate setup. Measure before you buy.
License note: a Territory-limited grant covers the EU, UK and South Korea exclusion, so confirm your company's location first.
Territory and the hosted alternative
Check the license before renting hardware: the Tencent Hunyuan license excludes the EU, UK and South Korea, and its Section 5(c) extends that to using or displaying outputs there.
Sume does not list HunyuanVideo. For a hosted equivalent in effort, Sume lists wan-3.0 at $0.125 per second at 720p (list $0.10 x 1.25), so a five-second 720p clip is $0.625. That is a different model, so judge the output, not just the price. The flow is a POST /v1/videos with a model id and a prompt, then a poll.
Sources
Related posts
More in Models
- HunyuanVideo 720p: 1,904 s on one GPU, 338 s on eight
README latency for a 1280x720, 129-frame, 50-step clip: 1,904 s on 1 GPU and 338 s on 8. The GPU-seconds each setup burns, and what it means for hosted jobs.
- Hy Image 3.5 lists a 1.5K tier. No Sume image model does
Hy Image 3.5 Preview offers 1K, 1.5K, 2K and 4K. Sume's tier lists are 512, 1K, 2K, 4K or 1K/2K only. For about 1.5K, GPT rows take 1536x1536 as image_size.
- Hy Image 3.5 Preview: 2K or 4K? Tencent and OpenRouter differ
Tencent's launch post says up to 2K; OpenRouter's page lists resolution 1K to 4K. How to test it, and which Sume image rows list a 4K tier or pixel size.
- Hy Image 3.5 lists 13 ratios; GPT Image 2.5 on Sume lacks six
OpenRouter's Hy Image 3.5 Preview ratio list has six ratios GPT Image 2.5 does not offer on Sume: 1:2, 1:4, 2:1, 2:3, 3:2 and 4:1. Which Sume rows list them.
Written by Sume