Cog
CogVideo System Requirements: VRAM, GPU & Hardware Guide
Everything you need to run CogVideoX locally: exact VRAM, GPU and system requirements for CogVideoX-2B, 5B and 5B-I2V. No suitable GPU? You can generate video online for free instead.
Quick answer
Minimum ~4 GB VRAM (CogVideoX-2B, optimized) · recommended 12 GB (RTX 3060) · or run it free online with no GPU.
CogVideoX VRAM requirements by model
Single-GPU inference VRAM with the diffusers optimizations enabled (enable_model_cpu_offload, vae.enable_tiling, etc.). Disabling them increases peak VRAM by roughly 3×.
| Model | Inference (optimized) | INT8 quantized | No optimization | Fine-tune (LoRA) |
|---|---|---|---|---|
| CogVideoX-2B | from 4 GB (FP16) | from 3.6 GB (INT8) | ~3× higher (≈12 GB+) | 47 GB (LoRA) |
| CogVideoX-5B | from 5 GB (BF16) | from 4.4 GB (INT8) | ~3× higher (≈15 GB+) | 63 GB (LoRA) |
| CogVideoX-5B-I2V (Image-to-Video) | from 5 GB (BF16) | from 4.4 GB (INT8) | ~3× higher (≈15 GB+) | 63 GB (LoRA) |
Fine-tuning figures apply to NVIDIA A100/H100. CogVideoX1.5-5B targets higher resolution and longer clips, so expect proportionally higher VRAM.
Recommended GPUs
Minimum · 6-8 GB VRAM
e.g. RTX 2060, RTX 3050, GTX 1660
Works with INT8 quantization + CPU offload enabled. Generation is slow but functional for CogVideoX-2B/5B.
Recommended · 12 GB VRAM
e.g. RTX 3060 12GB, RTX 4070
Comfortably runs both CogVideoX-2B and 5B with optimizations on. Best price-to-performance for local use.
High performance · 24 GB VRAM
e.g. RTX 3090, RTX 4090, A5000
Run without memory optimizations for the fastest inference and higher-resolution or longer clips.
Fine-tuning · 48-80 GB VRAM
e.g. A100, H100
Required for LoRA fine-tuning (47 GB for 2B, 63 GB for 5B at batch size 1).
Local software & system requirements
- Operating system: Linux (Ubuntu 20.04+), or Windows 10/11 (WSL2 recommended).
- Python 3.10 or higher (Anaconda recommended for environment setup).
- PyTorch 2.x built with CUDA 12.x support.
- diffusers >= 0.30.3 (only the diffusers build supports INT8/FP8 quantization).
- NVIDIA GPU with up-to-date drivers. With quantization it can fit on a free Colab T4.
- System RAM: 16 GB minimum, 32 GB recommended.
- Disk space: ~20-40 GB for model weights and dependencies.
New to local setup? Follow the CogVideo how-to-use guide for installation steps, or read the official GitHub documentation.
Online requirements (no GPU needed)
- A modern browser: Chrome, Firefox, Edge or Safari (latest version).
- A stable internet connection (10 Mbps+ for smooth generation).
- A free CogVideo account. No GPU, drivers or installation required.
Frequently asked questions
How much VRAM do I need to run CogVideo?
With diffusers memory optimizations enabled, CogVideoX-2B runs from about 4 GB of VRAM (3.6 GB with INT8) and CogVideoX-5B from about 5 GB (4.4 GB with INT8). Disabling those optimizations raises peak VRAM roughly 3 times. For a smooth experience we recommend a 12 GB GPU such as an RTX 3060.
Can I run CogVideoX on an RTX 3060?
Yes. A 12 GB RTX 3060 comfortably runs both CogVideoX-2B and CogVideoX-5B with the standard diffusers optimizations enabled. It is the recommended entry-level card for local CogVideoX generation.
Can I run CogVideo without a GPU?
Not at usable speeds locally, CogVideoX needs an NVIDIA GPU. If you don't have one, use CogVideo Online to generate Text-to-Video, Image-to-Video and Video-to-Video in your browser for free, with no GPU or setup.
What are the minimum system requirements for CogVideoX?
Python 3.10+, PyTorch 2.x with CUDA 12.x, diffusers >= 0.30.3, an NVIDIA GPU with at least 6-8 GB VRAM, 16 GB system RAM, and roughly 20-40 GB of free disk space for model weights.
What is CogVideo / CogVideoX?
CogVideo (and the newer CogVideoX family: 2B, 5B and 5B-I2V) is an open-source text-to-video and image-to-video generation model developed by THUDM/Zhipu AI. You can run it locally or use it free online here on CogVideo.