Beijing Lab Open-Sources 114B MoE Audio/Video Model MAGI-2

ShanRizvi · x · 2026-08-15

A new Beijing-based lab founded by Swin Transformer creators Cao Yue and Zhang Zheng has open-sourced MAGI-2 Preview. This is a 114B parameter Mixture-of-Experts (MoE) model for audio and video generation. It activates only 6B parameters per token via ultra-fine-grained head routing and custom MagiMoE kernels, supporting single-stream native A/V generation at practical scale. The generation cost is claimed to be 1/10th of mainstream models. Code and tech report are fully available.

Related event: SandAI Open-Sources MAGI-2, a 114B-Parameter MoE Video Model(4 posts)→

Original post →

More from Multimodal

Multimodal channel →