Tencent HunyuanVideo 13B (Cinema DiT) VRAM Requirements Guide
Cinema-grade 13B video synthesis Diffusion Transformer with 3D VAE. Ideal for 24GB RTX 3090/4090 (quantized) or 48GB+ rigs.
>98.2% retention. 10.7 GB weights, leaving +7.6 GB headroom on Dual NVIDIA RTX 3060 Rig (24GB VRAM).
Near-lossless 8-bit precision requiring 25.2 GB VRAM for benchmark evaluations.
Uncompressed FP16 weights taking 41.7 GB. Demands multi-GPU or datacenter accelerators.
Architectural Memory Breakdown & Attention Scaling
Tencent's Hunyuan Video 13B is a 13.0B parameter open-weights video diffusion transformer engineered for high-fidelity 720p generative video. Pairing dual text encoders (multilingual Hunyuan-CLIP and MLLM) with a 3D causal VAE, native weights and temporal latent buffers consume 21.5 GB in 4-bit precision, serving as the benchmark video workload for 24GB GPUs.
In ComfyUI or the native Hunyuan pipeline, pass `--fp8_e4m3fn` for text encoders and enable temporal VAE tiling. This prevents out-of-memory spikes during the final 3D latent decompression pass on 24GB graphics cards.
Dual NVIDIA RTX 3060 Rig (24GB VRAM) provides 24.0 GB physical VRAM (+7.6 GB operational margin at 8k context).
Run Tencent HunyuanVideo 13B (Cinema DiT) Locally
ollama run hunyuan-video-13b:q4_K_Mvllm serve Tencent/hunyuan-video-13b --max-model-len 8192 --gpu-memory-utilization 0.95Frequently Asked Questions
What GPU is required to run Hunyuan Video 13B locally?
An RTX 3090 (24GB) or RTX 4090 (24GB) is the minimum practical GPU for local generation. GPUs with 16GB VRAM will trigger allocation crashes.
Does Hunyuan Video support image-to-video conditioning?
Yes! Hunyuan Video supports both text-to-video and image-to-video generation, encoding starting frames directly into 3D latent video volumes.
Test Custom Hardware & Context Windows
Simulate Tencent HunyuanVideo 13B (Cinema DiT) across all 31 GPUs, adjust context up to 512 tokens, or calculate RunPod break-even costs.
Open Tencent HunyuanVideo 13B (Cinema DiT) in GPU Garden Matchmaker →