project
MAGI-2-preview - Sand.ai's open-source multimodal video generation model
MAGI-2-preview is Sand.ai's open-source MoE multimodal video generation model, boasting a total of 114 bytes of parameters and only 6 bytes of activation. The model continues the single-stream architecture, integrating text, video, and audio into a single T...
What is MAGI-2-preview?
MAGI-2-preview is Sand.ai's open-source MoE multimodal video generation model, boasting 114B parameters and only 6B activations. Continuing the single-stream architecture, the model integrates text, video, and audio into a single Transformer for joint modeling, achieving native multimodal generation. It ranks sixth on the AA video generation leaderboard. The model code and pre-trained weights are available, validating a new and efficient scaling approach in the video generation field.
Main functions of MAGI-2-preview
- Multimodal video generationIt supports the generation of text, images, and videos, and natively integrates audio understanding.
- MoE Sparse ActivationOnly 6B of the 114B total parameters are activated, significantly reducing inference costs.
- Single-stream unified architectureIt abandons the traditional cross-attention stitching method and models both the image and sound from the first layer.
- High-efficiency ScalingTo address the communication and training stability issues in video generation scenarios, MoE was reconstructed to resolve cross-card communication bottlenecks under ultra-long sequences.
The technical principles of MAGI-2-preview
- MoE Hybrid Expert ArchitectureThe model has a total capacity of 114 bytes, and only 6 bytes of parameters are activated in a single forward propagation, thus decoupling capacity from computation.
- Single-stream TransformerText, video, and audio tokens enter the same Transformer from the input layer and exchange cross-modal information directly through self-attention, which is more conducive to capturing audio-visual synchronization details compared to multi-tower splicing architecture.
- Distributed communication optimizationTo address the characteristics of extremely long video sequences, the token routing and communication strategies in expert parallel processing have been restructured to reduce cross-GPU data transfer overhead.
- Training stability schemeA triple stability guarantee mechanism consisting of dynamic routing, expert load balancing, and multimodal data joint training.
Follow us on WeChat and reply with "open source",join inAI open source project discussion group
How to use MAGI-2-preview
- Environmental preparationCloning a GitHub repository
SandAI-org/MAGI-2-previewConfigure dependencies according to the README. - Download weightGet pre-trained weights from GitHub Releases.
- Running inference: Modify configuration files and script parameters to support
t2v/i2v/v2vThree modes, referencing MAGI-1run.shFormat execution. - Hardware RequirementsSpecific configurations will be updated in the official documentation. It is recommended to refer to the MAGI-1 24B multi-card H100 or 4.5B single-card 24GB VRAM solutions.
The core advantages of MAGI-2-preview
- Open source and commercially usableThe model is licensed under the Apache 2.0 license, and the pre-trained weights and inference code are completely open.
- MoE, a pioneering initiative with a scale of hundreds of billions.The world's first successfully open-sourced MoE video generation model with a capacity of hundreds of billions of videos validates a new path for scaling in the video field.
- Low cost and high capacityA 114B model can be driven with only 6B activation parameters, and the inference cost is much lower than that of a dense model of the same size.
- Native MultimodalThe single-stream architecture allows for deep fusion of audio and video at the underlying model level, avoiding latency and quality loss caused by later alignment.
The project address for MAGI-2-preview
- Project official 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 Dimensions | MAGI-2-preview | Wan2.7-Video |
|---|---|---|
| Development Team | Sand.ai | Ali Tongyi Lab |
| Architecture | MoE + Single-stream Autoregressive Transformer | 3D DiT + MoE + Flow Matching |
| Total parameters | 114B | 270B (14B series) |
| Activation parameters | 6B | 140B |
| Generative paradigm | Since the return | Diffusion (flow matching) |
| Multimodal | Native audio and video joint modeling (single-stream unified) | Full-modal input of text/image/video/audio |
| Video length | Unclear (continuing with MAGI-1 streaming capabilities) | 2-15 seconds |
| resolution | Unclear | 720P / 1080P |
| Open source license | Apache 2.0 (commercially usable) | Open source (weight open) |
| Core competencies | High-efficiency Scaling, Native Multimodal Video Generation | Video editing, reference video, camera movement control, facial expression driving |
| Ranking Performance | AA ranked sixth on the video generation list | DesignArena V2V Top Ranking |
| Editing ability | Unclear | One-line video editing, style transfer, partial replacement |
Application scenarios of MAGI-2-preview
-
Long-form video ads and short dramasMoE offers high-capacity support for complex narrative modeling and can generate minute-long, coherent story videos.
-
Multimodal film and television pre-visualizationNative audio understanding capabilities enable the simultaneous generation and editing of dubbing, sound effects, and visuals.
-
AI Video Academic ResearchThe open-source MoE weights for hundreds of billions of videos provide a reproducible scaling baseline for academia.
-
Enterprise privatization deploymentThe model is based on the Apache 2.0 license and supports the construction of private video generation capabilities in sensitive industries such as finance and healthcare.
-
Real-time streaming generationSparse activation reduces inference costs and can support low-latency real-time video generation and live streaming scenarios.