Video generation model

Self-hosting HunyuanVideo: 8.3B parameters, and one 24GB card gets you rendering

Tencent cut the original 13B Hunyuan video model down to 8.3B and squeezed 480p image-to-video into 8 steps. Here are the real VRAM figures, real step counts and real timings for every variant — plus which card to actually rent.

HunyuanVideo · self-hosted

Search "HunyuanVideo VRAM requirements" and you will still land on the table from late 2024: 60GB for 720×1280 at 129 frames, 45GB for 544×960, with an 80GB card recommended outright. That table is accurate — it just describes the original 13B model. In November 2025 Tencent Hunyuan shipped HunyuanVideo 1.5, dropping from 13B to 8.3B parameters, with an official floor of 14GB VRAM when model offloading is enabled. ComfyUI puts it more bluntly: the target is flagship-quality video on a 24GB consumer GPU. If you have been saving up for an A100 just to run Hunyuan video, read this page first.

The 1.5 speedup did not come from gutting quality. It ships a new 3D causal VAE with 16× spatial and 4× temporal compression, which slashes the token count reaching the DiT. Attention was replaced with SSTA — selective and sliding tile attention — which prunes redundant spatiotemporal KV blocks; Tencent measured a 1.87× end-to-end speedup over FlashAttention-3 on 10-second 720p synthesis across 8 H800s. On the text side it pairs Qwen2.5-VL-7B for semantics with a byT5-small GlyphXL encoder for glyph-aware conditioning, so Chinese and English prompts both land and on-screen text holds together far better than in the first generation. A separate super-resolution network takes 720p output up to 1080p instead of paying 1080p memory costs during generation.

The variant sprawl is both the project's strength and its trap. The repo carries 480p and 720p T2V and I2V, CFG-distilled builds, a 480p I2V step-distilled build, a 720p I2V sparse-attention build, plus 720p and 1080p super-resolution weights — 372GB in total on Hugging Face. Cloning it blindly will fill your disk and your afternoon. Weights ship under the Tencent Hunyuan Community License: commercial use is allowed, but the agreement explicitly does not apply in the EU, UK or South Korea, and anything past 100 million monthly active users needs a separate licence from Tencent. What actually blocks most people is still the card. The model is small enough for consumer hardware now, but the 8GB in your laptop still will not open the door. That is the whole case for renting a 4090 or 5090 by the second.

01 —

Which HunyuanVideo variant, and what it costs in VRAM

Every parameter count, step count and memory figure below comes from the official repo and model cards

Version	Parameters	VRAM	Context	Notes
HunyuanVideo 1.5 · 720p T2V / I2V	8.3B DiT + 3D causal VAE	14GB official floor (with offloading) / 24GB comfortable / 32GB to run without offload	720p, 121 frames by default, 5–10 seconds	The current mainline. cfg 6, 50 steps; flow shift 9 for T2V, 7 for I2V. Highest quality tier — render at 720p, then upscale.
HunyuanVideo 1.5 · 480p I2V step-distilled	8.3B (step-distilled)	Runs on a single 24GB card — the official benchmark card is the RTX 4090	480p, 8–12 steps, flow shift 7	About 75% off end-to-end generation time; roughly 75 seconds per clip on one RTX 4090. The right build for prompt sweeps and storyboard drafts. Step distillation covers 480p I2V only.
HunyuanVideo 1.5 · CFG-distilled (480p / 720p)	8.3B (cfg-distilled)	Same as the non-distilled build at the same resolution	480p / 720p, cfg set to 1	Embedded CFG drops the negative branch for roughly 2× speedup. Caveat: Tencent states plainly that CFG-distilled models must still run the full 50 steps — cut to 8 and you get garbage.
HunyuanVideo 1.5 · 1080p SR step-distilled	Standalone super-resolution network	An extra pass on the same card that rendered 720p; 32GB is roomier	720p → 1080p, 6–8 steps, flow shift 2	The official route to 1080p. Cheaper in both memory and time than generating 1080p directly. A 720p-tier SR checkpoint also exists.
HunyuanVideo (original 13B, Dec 2024)	13B DiT, Llava-Llama3-8B + CLIP dual text encoders	60GB at 720×1280×129f; 45GB at 544×960×129f; 80GB recommended	540p / 720p, 129 frames (frame count must be 4k+1)	The generation with the thickest ecosystem — LoRAs, I2V, Avatar and Foley all grew around it. --use-fp8 saves roughly 10GB. Pick it if you need existing LoRAs or style fine-tunes.
Original 13B quantized (GGUF / NF4)	13B quantized	Q8_0 weights 14GB / Q6_K 11GB / Q4_K_M 7.88GB; bnb NF4 in diffusers is ~14GB for the whole pipeline	Start at 61 frames and 320×512 when testing	city96's GGUFs pair with the ComfyUI-GGUF nodes; original BF16 weights are 25.6GB. You trade quality for memory — detail and motion consistency degrade at low bit widths.

02 —

Which GPU to rent for HunyuanVideo

Match the card to the job — an 8.3B model does not need H100 money

Batch 480p I2V step-distilled clips, iterate on prompts
RTX 4090 24GB$0.540/GPU-hour
The official 75-second benchmark was measured on an RTX 4090, so you get the published number on the published card with no extrapolation.
720p T2V/I2V at the full 50 steps with offloading off, then upscale to 1080p
RTX 5090 32GB$0.723/GPU-hour
The extra 8GB is exactly what lets you disable offloading — what you save is not memory, it is shuttling weights between CPU and GPU on every step.
Original 13B in bf16 at 720×1280×129 frames, or LoRA fine-tuning
A100 SXM4 80GB$1.088/GPU-hour
Tencent states 60GB is required for 720p at 129 frames. A 24GB card can only reach this tier through quantization compromises; 80GB is the answer that needs no compromise.
SSTA sparse attention, multi-GPU parallel inference, 10-second shots
H100 SXM 80GB$3.582/GPU-hour
The flex-block-attn sparse attention path requires H-series hardware, and the official 1.87× speedup was measured on an H800 cluster.

03 —

Deploying HunyuanVideo 1.5 in four steps

From cold instance to first rendered clip, this is the whole path

01
Boot a CUDA instance
Pick an RTX 4090 or RTX 5090 in the NexGPU console and choose a prebuilt PyTorch image — CUDA, drivers and torch are already in place. SSH in, confirm the card is visible, then install the repo's dependencies. Tencent requires Python 3.10+ and recommends FlashAttention.
```
nvidia-smi && git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5 && cd HunyuanVideo-1.5 && pip install -r requirements.txt
```
02
Download only the variant you need
The full tencent/HunyuanVideo-1.5 repo on Hugging Face is 372GB because it carries every resolution and every distilled build. Use --include to pull precisely what you want: one 720p transformer plus the VAE and SR network is a few dozen GB and enough to start. Inbound traffic is not billed, so the download itself costs nothing beyond the machine-hours it occupies.
```
huggingface-cli download tencent/HunyuanVideo-1.5 --local-dir ./ckpts --include "transformer/720p_i2v/*" "vae/*"
```
03
Set the memory knobs before you generate
If you hit OOM, the first move is the allocator environment variable — it is called out by name in the official README. If host RAM is tight, also turn off overlapped group offloading. Remember to drop the step-distilled 480p build to 8–12 steps, while the CFG-distilled build must stay at 50. Getting those two backwards produces either garbage frames or a wasted hour.
```
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:128
```
04
Render, then decide what to keep
Set nproc_per_node to 1 for single-card runs. One more thing: prompt rewriting is enabled by default and expects a separate vLLM server hosting a Qwen3-235B-class model — on a single card, switch it off and write detailed prompts yourself. Pull your clips down when you are finished; compute billing stops when the instance stops, while storage keeps accruing until the volume is destroyed.
```
torchrun --nproc_per_node=1 generate.py --prompt "An orange cat leaps onto a windowsill, sunlight through sheer curtains" --resolution 480p --model_path ./ckpts --video_length 121 --enable_step_distill true
```

What a batch of HunyuanVideo clips actually costs

Take the common path: RTX 4090 24GB at $0.540/GPU-hour, running 480p I2V step-distilled. Spend roughly 0.4 hours installing dependencies, pulling weights and warming the model — 0.4 × $0.540 = $0.216. The official benchmark is 75 seconds per clip; budget 90 seconds with VAE decode and disk writes, so 3600 ÷ 90 = 40 clips per hour. Run two hours for 80 clips: 2 × $0.540 = $1.08. Total $1.296, or about $0.016 per clip. Move up to 720p at 50 steps on an RTX 5090 32GB at $0.723/GPU-hour and the hourly rate is 34% higher, but you drop the offload shuffling, so per-clip cost does not shift by an order of magnitude. The number worth watching is storage: at the $0.414/GB-month median, keeping 30GB of weights on disk is $12.42 a month — over ten times what those two hours of compute cost. Destroy the volume when you are done and re-pull next time; inbound transfer is free, and dragging 80 finished 480p clips home is roughly 1GB, which at the $0.0081/GB median egress rate is under a cent. NexGPU meters per second and prices per hour, with no minimum, no setup fee and no quota request — stop the instance and compute charges stop with it.

04 —

Frequently asked questions

How much VRAM does HunyuanVideo actually need to run locally?

It depends which generation you mean. HunyuanVideo 1.5 (8.3B) has an official floor of 14GB with model offloading enabled, though the comfortable zone is 24GB — which is also the number ComfyUI markets it on. The original 13B is far heavier: 45GB for 544×960 at 129 frames and 60GB for 720×1280 at 129 frames, with an 80GB card recommended. The shortcut is to rent the tier you actually need — on NexGPU an RTX 4090 24GB is $0.540/GPU-hour and an A100 SXM4 80GB is $1.088/GPU-hour, billed per second, so guessing wrong just means stopping and switching cards.

Can a single 24GB RTX 4090 run HunyuanVideo 1.5?

Yes, and the 4090 is the official benchmark card. The 480p I2V step-distilled build renders a clip in about 75 seconds on one RTX 4090, roughly 75% faster end-to-end than the undistilled path. The 720p tier also runs, but 24GB means offloading weights, which costs you speed; if you want offloading off, 32GB is the easier answer. NexGPU lists the RTX 4090 at $0.540/GPU-hour and the RTX 5090 32GB at $0.723/GPU-hour, both with PyTorch and ComfyUI prebuilt images so you can start pulling weights immediately.

HunyuanVideo 1.5 or the original 13B — which should I pick?

For anything new, take 1.5: parameters drop from 13B to 8.3B, the VAE hits 16× spatial and 4× temporal compression, SSTA attention measured 1.87× faster than FlashAttention-3 on 10-second 720p across 8 H800s, and you get 1080p super-resolution plus bilingual glyph-aware encoding. There is one good reason to stay on the original: you depend on the large body of existing 13B LoRAs, or you need the older HunyuanVideo-I2V, Avatar and Foley ecosystem. If you want to try both, no need to agonize — swapping cards and machines on NexGPU takes minutes, with nothing locked behind a monthly commitment.

Why does the CFG-distilled build produce broken video when I lower the step count?

This is HunyuanVideo's nastiest gotcha, and the official README says it outright: CFG-distilled models must run the full 50 steps to produce correct results. What was distilled away is the negative branch (cfg set to 1), worth roughly 2× — it is not a step-count speedup. The build that genuinely takes fewer steps is the 480p I2V step-distilled one at 8–12 steps with flow shift 7, plus the SR network at 6–8 steps. Confusing the two is the single most common way newcomers burn an hour of 4090 time for nothing. On NexGPU that hour is $0.540 — cheap, but the time does not come back.

Can I use HunyuanVideo commercially, and what does the licence restrict?

It ships under the Tencent Hunyuan Community License, which permits commercial use on a non-exclusive, non-transferable, royalty-free basis. Two hard limits are worth memorizing: the agreement explicitly does not apply in the European Union, the United Kingdom or South Korea, and any product past 100 million monthly active users must request a separate licence from Tencent. Beyond that, self-hosting carries no call quotas or approval steps. Same on our side: no quota request, no minimum commitment, and 1,175 verified rentable nodes across 51 countries and regions to place the workload wherever you need it.

Do I really need to deploy a separate 235B model just for prompt rewriting?

It is on by default, and the recommended rewriters are Qwen3-235B-A22B-Thinking-2507 for text-to-video and Qwen3-VL-235B-A22B-Instruct for image-to-video, each expecting its own vLLM server. On a single card that is not worth it — switch it off and write detailed prompts yourself, covering subject, action, lighting and camera movement, and you land in the same place. If you do want to build the two-service pipeline properly, NexGPU supports up to 14 GPUs per node and a maximum node VRAM of 2,152GB, so the rewriter and the video model can share one machine over the internal network. Support is bilingual over Telegram, with no ticket queue.

Start building on NexGPU

Enterprise R&D team or solo developer — either way, your first job can be running in minutes.