MAGI-2-preview - Sand.ai's Open-Source Multimodal Video Generation Model

What is MAGI-2-preview

MAGI-2-preview is the world's first 100-billion-parameter MoE multimodalvideo generationmodel open-sourced by Sand.ai, with 114B total parameters and only 6B activated. The model continues the single-stream architecture, unifying text, video, and audio into the same Transformer for joint modeling, achieving native multimodal generation and ranking sixth on the AA video generation leaderboard. The model code and pre-trained weights have been released, validating a new path for efficient Scaling in the video generation field.

Main Features of MAGI-2-preview

Multimodal video generation: Supports text, image, and video-to-video generation, with native integration of audio understanding.
MoE sparse activation: Only 6B of the 114B total parameters are activated, greatly reducing inference costs.
Single-stream unified architecture: Abandons the traditional cross-attention concatenation approach, modeling visuals and sound jointly from the very first layer.
Efficient Scaling: Reconstructs MoE communication and training stability for video generation scenarios, solving the cross-GPU communication bottleneck under ultra-long sequences.

Technical Principles of MAGI-2-preview

MoE mixture-of-experts architecture: The model has a total capacity of 114B, with only 6B parameters activated per forward pass, decoupling capacity from computation.
Single-stream Transformer: Text, video, and audio tokens enter the same Transformer from the input layer, directly exchanging cross-modal information through self-attention, which is more conducive to capturing audio-visual synchronization details than multi-tower concatenation architectures.
Distributed communication optimization: Targeting the ultra-long sequence characteristics of video, it reconstructs the token routing and communication strategies in expert parallelism, reducing cross-GPU data transfer overhead.
Training stability solution: A triple stability safeguard mechanism of dynamic routing, expert load balancing, and joint multimodal data training.

Follow WeChat and reply "open source" to join theAI Open-Source Project Exchange Group

How to use MAGI-2-preview

Environment preparation: Clone the GitHub repository SandAI-org/MAGI-2-preview, and configure dependencies according to the README.
Download weights: Obtain pre-trained weights from GitHub Releases.
Run inference: Modify the configuration file and script parameters; supports t2v / i2v / v2v three modes, refer to MAGI-1's run.sh format for execution.
Hardware requirements: Specific configurations are pending official documentation updates; it is recommended to refer to MAGI-1's 24B multi-GPU H100 or 4.5B single-GPU 24GB VRAM solution.

Core Advantages of MAGI-2-preview

Open source and commercially usable: The model uses the Apache 2.0 license, with pre-trained weights and inference code fully open.
First 100-billion-parameter MoE: The world's first successfully open-sourced 100-billion-level MoE video generation model, validating a new Scaling path in the video field.
Low cost, high capacity: 6B activated parameters are enough to drive a 114B model, with inference costs far lower than dense models of the same scale.
Native multimodal: The single-stream architecture enables deep fusion of audio and video at the model's底层, avoiding the latency and quality loss caused by post-hoc alignment.

MAGI-2-preview project address

Project website：https://sand.ai/blog/magi-2-preview
GitHub repository：https://github.com/SandAI-org/MAGI-2-preview
HuggingFace model library：https://huggingface.co/sand-ai/MAGI-2-preview

Comparison of MAGI-2-preview with similar competing products

Comparison dimension	MAGI-2-preview	Wan2.7-Video
Development team	Sand.ai	Alibaba Tongyi Lab
Architecture	MoE + single-stream autoregressive Transformer	3D DiT + MoE + flow matching
Total parameters	114B	270B (14B series)
Activated parameters	6B	140B
Generation paradigm	Autoregressive	Diffusion (flow matching)
Multimodal	Native joint audio-video modeling (single-stream unified)	Text/image/video/audio all-modal input
Video duration	Not specified (continues MAGI-1 streaming capability)	2-15 seconds
Resolution	Not specified	720P / 1080P
Open-source license	Apache 2.0 (commercially usable)	Open source (weights open)
Core capabilities	Efficient scaling, native multimodal video generation	Video editing, reference-to-video generation, camera movement control, expression driving
Leaderboard performance	No. 6 on the AA video generation leaderboard	No. 1 on DesignArena V2V
Editing capability	Not specified	One-sentence video modification, style transfer, local replacement

Application scenarios of MAGI-2-preview

Long-video advertising and short dramas: MoE's large capacity supports complex narrative modeling and can generate minute-level coherent story videos.
Multimodal film and television preview: Native audio understanding enables synchronized generation and editing of dubbing, sound effects, and visuals.
AI video academic research: Open-source hundred-billion-scale video MoE weights provide academia with a reproducible scaling baseline.
Enterprise private deployment: Based on the Apache 2.0 license, the model supports sensitive industries such as finance and healthcare in building private video generation capabilities.
Real-time streaming generation: Sparse activation reduces inference costs and can support low-latency real-time video generation and live streaming scenarios.