Skip to main content
llm.info

CogVideoX1.5-5B

Open Weights

Zhipu AI

Zhipu AI's open-weight text-to-video model, released 8 November 2024 as the newest generation of the CogVideoX family. It generates 1360x768 video at 16 frames per second in 5- or 10-second lengths, with frame counts of 16N+1 where N is at most 10 (default 81). Prompts must be English and are capped at 224 tokens. The vendor reports diffusers inference from 10 GB of VRAM in BF16 and from 7 GB with INT8, against 76 GB for the SAT BF16 path, and roughly 1000 seconds on a single A100 for a five-second clip. Image-to-video lives in a separate checkpoint, CogVideoX1.5-5B-I2V.

Strengths

  • Open weights with a first-class Diffusers integration through CogVideoXPipeline
  • 1360x768 output at 16 frames per second, up from 720x480 at 8 fps in CogVideoX-5B
  • Generates 5- or 10-second clips using 16N+1 frames with N up to 10
  • Runs from 10 GB VRAM in BF16 and from 7 GB with INT8 torchao quantisation via diffusers
  • Supports BF16, FP16, FP32, FP8 and INT8 inference precisions
  • Distributed on Hugging Face, ModelScope and WiseModel, with an SAT build for fine-tuning

Caveats

  • The CogVideoX License is not open source: commercial use requires registering for a basic commercial licence, and that licence stops at 1 million visits per month
  • English-only prompts capped at 224 tokens - the card says other languages "should be translated into English during prompt crafting with a larger model"
  • Slow by the vendor's own figures: about 1000 seconds on a single A100 and 550 seconds on a single H100 for a five-second video, and 76 GB for SAT BF16
  • Text-to-video only in this repository; image-to-video needs the separate CogVideoX1.5-5B-I2V checkpoint
  • The card publishes no video-quality benchmark scores at all, only memory and inference-speed figures
Vendor
Zhipu AI
Context Window
224 tokens
License
CogVideoX License
Release Date
2024-11-08
Modalities
video

Capabilities

Vision
Audio
Video
Tool Use

Pricing

Weights download free, but the CogVideoX License grants academic research use by default: "Users who wish to use the models for commercial purposes must register and obtain a basic commercial license". That basic commercial licence is free but capped, since "the number of service users (visits) for your commercial activities must not exceed 1 million visits per month"; above that a further commercial licence must be negotiated. Zhipu also sells larger hosted video models through QingYing and its API platform.

Reviews

Comments