Sand.ai Open-Sources 100B-Param Video Model MAGI-2 at 1/10th Inference Cost

量子位 · wechat · 2026-08-05

Tsinghua-backed team Sand.ai has released and open-sourced MAGI-2-preview, a 100-billion-parameter video generation model. Despite having 114B total parameters, it activates only about 6B parameters during a single forward pass through an extreme MoE architecture. Generating a 10-second 1080P video on 8x H100s costs only 0.5 RMB, roughly one-tenth the cost of mainstream industry models.

To overcome the communication and memory bottlenecks caused by ultra-long video sequences, the team built a custom infrastructure: a single-stream unified architecture for audio-visual processing, alongside a Multi-Head Latent MoE and Head Parallel strategy that decouples communication volume from the number of activated experts. The model ranks 6th on authoritative video generation leaderboards, closely trailing top-tier closed-source models.

Original post →

More from Multimodal

Multimodal channel →