On August 5, Sand.ai released and open-sourced MAGI-2 Preview, the world's first 100B+ MoE video generation model, with 114B total parameters and only 6B activated. The model uses an Ultra-fine-grained MoE architecture to unify text, video, and audio in a single Transformer, aiming to reduce generation costs and advance open-source video models.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
On August 5, Sand.ai released and open-sourced MAGI-2 Preview, a 114B-parameter model with 6B activated parameters, making it the world's first open-source 100B-scale MoE video generation model. Through fine-grained experts and Head Parallel architecture, the cost of generating a 10-second video is reduced to approximately 0.5 RMB, about one-tenth that of mainstream industry models.
Video generation models have long been limited to parameter sizes under 10B, with the largest around 20-30B, while language models have already reached hundreds of billions or even trillions of parameters. The root cause of this gap is that video sequences are far longer than text: a single frame corresponds to a large number of spatial patches, and when text and audio are added over several seconds, the token count far exceeds that of typical language tasks.
In dense models, every token must pass through all parameters, so increasing parameters proportionally raises per-token computation. When multiplied by extremely long sequences, training and inference costs become prohibitive. When OpenAI shut down Sora in March 2026, a major factor was that the architecture and inference costs made large-scale commercialization economically unviable.
Language models use MoE to decouple total capacity from per-token computation, but when video models adopt MoE, traditional expert parallelism requires routing before communication. The more activated experts and the longer the sequence, the greater the cross-card communication data, which can easily offset the computational savings from sparsity.
MAGI-2 Preview employs Ultra-fine-grained MoE: the 3,072-dimensional hidden representation is split into 12 256-dimensional subspaces (heads), each with 256 experts from which 6 are selected, resulting in 3,072 independent small experts per layer and 72 activated per token per layer. The entire model has approximately 114B total parameters and activates about 6B.
For reference, DeepSeek-V4-Pro-Preview has 384 routed experts per layer with 6 selected per token; Kimi K3 has 896 experts with 16 selected per token. Although expert definitions and tasks differ, the comparison shows that MAGI-2's routing granularity is finer, allowing motion, appearance, sound, and semantics to each seek suitable experts.
The model places text, video, and audio in the same context, with shared and dedicated experts coexisting, influencing each other from the start of generation, reducing interface and alignment issues between visuals and sound.
Thousands of fine-grained experts bring high-frequency routing, data rearrangement, and small matrix computations. Sand.ai developed the computation kernel library MagiMoE to merge routing, sorting, and expert computation; training is supported by the distributed optimizer MagiMuon, which uses Muon for main matrix parameters and AdamW for the rest, enabling stable scaling to 100B parameters.
According to Sand.ai's internal estimates, on 8 H100 GPUs at current monthly rental prices, generating a 10-second video costs about 0.5 RMB (for the distilled model), translating to about one-tenth the cost per second of mainstream industry models.
A tenfold cost reduction makes previously uneconomical uses like batch generation of assets and multi-version iteration potentially routine in production workflows.
On Artificial Analysis' image-to-video leaderboard, MAGI-2 Preview ranks 6th, and an open-source model with only 6B activated parameters has entered the top tier. Sand.ai acknowledges there is still a gap with the best models, but it demonstrates that architecture and system optimization can improve GPU utilization efficiency.
Open-sourcing allows research institutions to dissect fine-grained expert division of labor, and inference frameworks and chip vendors gain new public workloads; enterprises can deploy the model in their own environments to handle confidential or copyright-restricted material and continue training on specific characters and products.
Application companies can recalculate costs among external APIs, local deployment, and customized models. However, deploying a 114B model still requires hardware, communication, engineering, and operational expertise. Open-sourcing changes the boundary of participation, not the barrier to entry.
The language model trajectory shows that once open-source models enter a usable range, enterprises compare deployment costs, data control, customization space, and vendor risk, and closed-source vendors face pressure to lower prices and open capabilities. Video generation may form a similar landscape.
Future competition will extend from generation quality to price, data, and development tools, and performance ceilings alone will not determine winners. MAGI-2 Preview will not immediately rewrite rankings, but it provides a public sample at the 100B-parameter scale for more people to dissect and validate.
This article is primarily based on Sand.ai's official release information and DeepTech's reporting. Cost data (0.5 RMB/10 seconds) is Sand.ai's internal estimate and has not been independently verified; the leaderboard ranking comes from Artificial Analysis, but specific evaluation details are not provided. The reason for OpenAI's shutdown of Sora is attributed in the report and is not officially confirmed.
MAGI-2 Preview, through fine-grained MoE and Head Parallel architecture, is the first to push open-source video models to 100B parameters and reduce generation costs to one-tenth of the industry, potentially changing the competitive dimensions of the video generation industry.
Primary report
Primary source