The "DeepSeek moment" for video generation may have genuinely arrived.

On August 5, Tsinghua-backed AI startup Sand.ai released and fully open-sourced the hundred-billion-parameter video generation model MAGI-2 Preview. Founder Cao Yue is a co-first author of Swin Transformer and a long-established figure in computer vision. What they've delivered this time approaches closed-source first-tier quality in parameter scale and output, while driving generation costs down to one-tenth of mainstream models.

The numbers are striking: 114 billion (114B) total parameters, with roughly 6 billion parameters activated per forward pass; generating a 10-second 1080P video costs about 0.5 RMB (~$0.07) — while current mainstream closed-source video models (Sora, Kling, Runway Gen-3) charge roughly 5–8 RMB for the same specification. One-tenth the price, near first-tier quality, ranked sixth on the AA video generation leaderboard. Kai-Fu Lee saw it and immediately offered his assessment: "This is the DeepSeek of AI video generation."

Why Video Models Can't Just Copy LLM Scaling

MAGI-2's core breakthrough isn't parameter scale; it's proving that video generation scaling can't simply copy the LLM playbook — architecture has to be rebuilt from the ground up.

Over the past year-plus, many companies building video generation models have been tinkering with LLM architectures — replacing text tokens with video tokens and taking large language model Transformers directly to video generation. The result? Models got bigger, training and inference costs skyrocketed, but improvements in generation quality hit diminishing returns. Longer videos, higher resolutions, better temporal consistency — every increment required exponentially more compute. This is video generation's "impossible triangle": bigger models, longer videos, controllable costs — pick any two.

MAGI-2 breaks this triangle with MoE (Mixture of Experts) architecture. But it didn't simply port MoE from LLMs — they designed a Multi-Head LatentMoE purpose-built for video, with three key innovations:

First, unified single-stream Transformer. Text, video, and audio are processed by the same backbone network rather than three separate encoders stitched together. This means the model natively understands correspondences between text descriptions, video content, and audio, yielding better multimodal alignment.

Second, a custom MagiMoE kernel library. MoE inference efficiency has long been a problem — routing, expert parallelism, load balancing; any weak link makes MoE slower than dense models. Sand.ai didn't use off-the-shelf MoE implementations; they wrote kernels from scratch optimized specifically for video generation scenarios, pushing activation sparsity efficiency to its limits.

Third, specialized optimization for video temporal modeling. The biggest difference between video and text is the temporal dimension — frames must be coherent; no flickering, no jumping, no objects disappearing and reappearing. MAGI-2 adds specific temporal consistency constraints in the attention mechanism, generating relatively stable long videos without heavy post-processing.

These three innovations together deliver the "hundred-billion parameters, ultra-low inference cost" effect. 114B total parameters sounds intimidating, but only around 6B parameters activate per generation — meaning you get near-hundred-billion model quality at 6B-model cost. That's the power of MoE architecture, and it's why DeepSeek, Mistral, and others already drove costs down dramatically in LLMs, while in video, nobody truly got it right until now.

The Shockwave of an Open-Source Video Model

MAGI-2 being fully open source (model weights downloadable on both GitHub and HuggingFace) may have a bigger impact than the model's quality itself.

The first shockwave is pricing. Sora, Kling, and Runway are still priced on a luxury "per-second" logic — generating a minute of high-quality video costs hundreds of yuan. MAGI-2 drives costs to one-tenth, and open source means anyone can build on it, deploy their own video generation services. Over the next year, video generation API prices will likely collapse the same way LLM API prices did. Video generation going from "professional tool" to "infrastructure everyone can afford" may happen faster than anyone expected.

The second shockwave is ecosystem. Open-source models will spawn massive vertical applications: short-video creators using it to batch-generate footage, e-commerce platforms using it to produce product videos at scale, ad agencies using it for quick turnarounds, education companies using it for course videos. Just as Stable Diffusion's open-sourcing spawned hundreds of startups in image generation, the startup wave in video generation may be about to detonate, ignited by MAGI-2.

Of course, MAGI-2 is still a Preview release, ranked sixth on the AA leaderboard — it still lags behind closed-source first-tier models like Sora and Kling, particularly on hard problems like complex motion scenes, long-video consistency, and text rendering. But it proves one thing: video generation doesn't require closed source, doesn't require sky-high compute budgets, doesn't require monopoly; innovative architecture design and open source can still produce near-top-tier results.

That's the power of open source. Llama going open source fundamentally changed the competitive landscape for foundation models. Will MAGI-2 do the same for video generation? We'll find out soon.

See you tomorrow.