Sand.ai just dropped MAGI-2 Preview, an open-source video model that generates 10-second clips with synced audio.
This model uses a clever fine-grained Mixture-of-Experts (MoE) architecture. Even though it has 114 billion parameters in total, only 6 billion are active for each step, making it super efficient for video generation.
Thanks to this MoE trick, inference cost is about 10x lower than dense models like MiniMax H3 (33B params), and it landed 6th place on the Artificial Analysis leaderboard. The generation pipeline is two-stage: a base model for low-res video, then a refiner upscales to 1080p. While MiniMax H3 fits on a single 24GB GPU, MAGI-2 needs more. But for cloud users, those 6B active parameters mean much cheaper API calls and good scaling potential.
Paper
GitHub
HuggingFace
#GenAI #AITools
September 1, 2026 630 1
