Video generation models are running into a fundamental problem: they're becoming too expensive to train and run, even as they get smarter. A newly open-sourced model called MAGI-2 Preview, developed by Sand.ai, offers a potential solution by using a technique borrowed from large language models that could reshape how video AI evolves over the next few years.

Why Are Video Models Hitting a Scaling Limit?

Two years ago, AI-generated videos were rough, short, and full of obvious flaws. Characters would distort, physics would break down, and waiting 30 minutes for a few seconds of usable footage was normal. Today, models from OpenAI's Sora and Google's Veo to China's Kling and MiniMax Hailuo can generate dozens of seconds or even several minutes of high-quality video with consistent characters and complex camera movements.

But this rapid progress has created an unexpected problem. Video generation is the most computationally intensive task in artificial intelligence. Unlike text, which is one-dimensional, video requires intense processing across both space (the image itself) and time (frame-to-frame continuity). When models add audio, dialogue, ambient sound, and background music on top of that, the amount of information explodes exponentially.

Traditional video models process every single parameter in the model for every piece of data they handle. As models grow larger and videos become more complex, the computing power required grows out of control. The industry is hitting what researchers call "the dual walls of physics and business," where the cost of training and running these models is approaching what's economically feasible.

What Is the MoE Solution, and Why Does It Matter?

Large language models faced a similar problem a few years ago. Their solution was a technique called Mixture of Experts, or MoE. Instead of activating every parameter in the model for every task, MoE builds an enormous library of specialized "expert" modules but only activates a small subset of them for each specific job. This clever approach lets models grow much larger without proportionally increasing computing costs.

MAGI-2 Preview demonstrates this approach applied to video. The model has approximately 114 billion total parameters, but only activates about 6 billion of them during each video generation task. According to Sand.ai's calculations, generating a 1080p video costs roughly 0.5 yuan (about 7 cents USD) per 10 seconds of footage. For a unified audio and video model of this size, that cost efficiency is remarkable because it doesn't scale linearly as the model expands.

The model currently ranks 6th globally on the Image to Video track of Artificial Analysis, a leading independent benchmark for video generation models.

Why Haven't Video Models Adopted MoE Sooner?

The answer lies in the fundamental differences between text and video. Text models process relatively small amounts of data; a few hundred words become a few hundred tokens after compression. A high-definition video longer than 10 seconds, even after heavy compression, generates tens of thousands or even hundreds of thousands of tokens. Video tokens also carry far more complex relationships than text tokens do.

Video tokens must handle multiple challenges simultaneously:

• Spatial smoothing: Ensuring pixels flow naturally across the image without jumps or artifacts
• Temporal continuity: Maintaining consistency across frames so objects move realistically over time
• Audio-visual synchronization: Aligning dialogue, lip movements, and ambient sound at millisecond precision
• Cross-modal coherence: Keeping dialogue, background music, and visual action in sync throughout the entire video

When millions of tokens with these complex interdependencies flood into an MoE system, engineering challenges emerge that text models don't face. The cross-frame attention mechanisms that video models rely on become significantly more complicated to implement efficiently.

How to Understand Video Generation's Next Evolution

• Parameter efficiency: Models can now grow much larger without proportional increases in computing cost, allowing for more sophisticated video generation without breaking budgets
• Unified audio-video generation: The industry standard is shifting toward models that handle dialogue, sound effects, and background music alongside video, requiring new architectural approaches
• Open-source acceleration: MAGI-2 Preview's full open-sourcing of model weights, code, and training systems means researchers worldwide can build on this approach rather than waiting for proprietary releases
• Scaling law revival: MoE architecture reopens the possibility of continued improvement through model scaling, even as dense model scaling hits its limits

The release of MAGI-2 Preview signals that video generation is entering what researchers call "the next Scaling Law era." For the past two years, the industry improved video quality primarily by making models bigger and training them longer. That approach is hitting diminishing returns. MoE offers a different path forward: smarter architecture that lets models grow without proportional cost increases.

This shift mirrors what happened in large language models. Models like Mixtral, DeepSeek, and Qwen gradually adopted MoE architecture as the standard approach, allowing them to scale beyond what dense models could achieve. Video generation appears to be following the same trajectory, though the engineering challenges are significantly steeper because of video's inherent complexity.

For creators, studios, and companies relying on video generation tools, this development matters because it suggests that future models will be both more capable and more affordable to run. The cost per second of generated video should continue declining as the industry refines MoE techniques specifically for video. The open-sourcing of MAGI-2 Preview also means that smaller teams and independent developers will have access to cutting-edge video generation capabilities, not just companies with massive computing budgets.