Sand.ai Open-Sources 114B MoE Model MAGI-2, Slashing Audio-Video Generation Costs by 90%

新智元 · wechat · 2026-08-06

Sand.ai has officially released and open-sourced MAGI-2 Preview, the world's first 100-billion-parameter MoE unified audio-video generation model. It has a total of 114B parameters but only activates about 6B per forward pass, reducing inference costs to 1/10th of mainstream models.

Core Architecture & Technical Breakthroughs:

Ultra-fine-grained MoE: Splits the 3072-dim representation into 12 256-dim subspaces. Each MoE layer has 3072 head-local expert units, but a token only activates 72, enabling richer capability combinations.
HeadParallel Communication: Flips the traditional "route-then-communicate" to "communicate-then-route". Cross-device transmission becomes fixed, eliminating the linear scaling of communication overhead with activated experts and solving the routing bottleneck.
Unified Audio-Video Generation: Drops the post-production stitching of video and audio. Text, video, and audio are modeled in the same context from step one, achieving high-fidelity audio-visual sync and lip alignment natively.

Performance & Cost:

Ranked #6 on the ArtificialAnalysis image-to-video leaderboard. Generating a 10-second video costs only $0.07 (0.5 RMB) based on 8x H100 rental rates, drastically lowering the commercial barrier for video products.

Original post →

Sand.ai Open-Sources 114B MoE Model MAGI-2, Slashing Audio-Video Generation Costs by 90%

More from Multimodal