Zangwei Zheng

Video-Ocean Video Agent generates videos up to minutes with a few clicks, including voice and face.

The world’s first open-source Sora-like video generation model — bringing efficient, high-quality video production to everyone.

Video-Ocean is a video generation platform that allows you to generate videos, images and audios with state-of-the-art models.

An optimizer that can train CTR prediction models with large batch (~128k)

We introduce CAME, a confidence-guided optimizer that combines the fast convergence of adaptive methods with the low memory footprint of memory-efficient ones.

MERIT constructs element-wise trust ratios to enable more robust update scaling by focusing on local weight structures.

A novel, elegant and super efficient sequence parallelism for multi-dimensional transformer architecture.

InfoBatch accelerates model training by pruning easy samples and reweighting losses, achieving up to 40% faster training without sacrificing accuracy across diverse tasks.

OpenMoE is a project aimed at igniting the open-source MoE community! We are releasing a family of open-sourced Mixture-of-Experts (MoE) Large Language Models.