Sand.ai opens first 100B-parameter MoE video model, slashing generation costs
Sand.ai opens first 100B-parameter MoE video model, slashing generation costs
Tsarist-era Chinese startup Sand.ai has released MAGI-2-preview, an open-source mixture-of-experts (MoE) video generation model with 114 billion total parameters but only around 6 billion activated per forward pass. The company claims this makes it the first 100B-parameter MoE video model in the world, and it ranks sixth on the AA video generation benchmark despite using just a fraction of its total parameters. Generating a 10-second 1080p video reportedly costs about 0.5 yuan (roughly 7 cents), approximately one-tenth the cost of leading closed models. The model builds on Sand.ai's previous single-stream architecture, which jointly models text, video, and audio from the first transformer layer, and incorporates a custom "MagiMoE" kernel library and a novel parallel strategy that decouples communication volume from expert activation counts, addressing a key scaling bottleneck for long video sequences.
This release represents a significant structural move in the video generation market. By pairing 100B-parameter scale with MoE efficiency and open weights, Sand.ai is attempting to write a new playbook for how video models scale — one that diverges from the dense, token-heavy approach used by most LLMs. The company's founder, Cao Yue (曹越), is a Tsinghua University standout who previously co-founded Guangnian Zhichu (光年之外) and led multimodal research at the Beijing Academy of Artificial Intelligence. Kai-Fu Lee has publicly compared the team to "the DeepSeek of AI video generation." If the model's quality holds up in real-world usage, it could pressure closed-source rivals like Kuaishou's Kling or ByteDance's Jimeng to justify their pricing, and give researchers and enterprises a viable self-hosted alternative for sensitive or high-volume video generation tasks.
From a market perspective, this is a template for how open-source video AI tries to outmaneuver closed systems: not by matching them on raw quality, but by weaponizing cost efficiency and extensibility. The 6B activation parameter count is notable because it implies that a 100B-parameter model can be run on a single high-end GPU at inference time, making private deployment far more accessible than dense models of comparable size. However, the company is careful to note that the 6B figure does not fully account for communication overhead, routing, or memory usage in practice. The move also reinforces a broader trend in the substrate: open-weight models increasingly challenge closed ones on performance-per-dollar, and infrastructure choice becomes a competitive moat. Whether this translates into sustainable developer adoption or meaningful enterprise traction remains to be seen, but Sand.ai has now placed itself as a leading contender in the open-source video generation arena.
