Open MoE Video Model MAGI-2 Preview by Sand.ai
Developers from Sand.ai have released the weights and inference code for the MAGI-2 Preview model, which generates video with synchronized audio.
Open MoE Video Model MAGI-2 Preview by Sand.ai
Developers from Sand.ai have released the weights and inference code for the MAGI-2 Preview model. The model is built on a single-stream architecture and generates 10-second clips along with synchronized audio tracks.
It employs a rare fine-grained MoE for such models. With a total of 114B parameters, only 6B are activated for each token. In terms of computation per step, it is significantly lighter than the dense MiniMax H3 (33B), which has reduced inference costs by about 10 times, and the model has climbed to 6th place in the Artificial Analysis leaderboard.
The generation pipeline is two-stage. First, the base model generates video at a low resolution, and then a refiner upscales the result to 1080p.
Currently, it is not a competitor to MiniMax H3 in terms of local availability. The quantized H3 fits on a single 24GB card (RTX 3090/4090), while MAGI-2 requires significantly more resources. However, for cloud providers, the inference via API will be much cheaper due to the only 6B active parameters, and this architecture has good scaling potential.
Why it matters
AnalysisThis model opens up new possibilities for generating video with synchronized audio, significantly reducing inference costs. Its innovative architecture also has potential for further development and scaling.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
GPT Image Generative Models Prompting Guide
OpenAI's gpt-image generative models are designed to produce production-quality visuals and fully controllable creative workflows. They are suitable for both professional design ta...
Claude Sonnet 4.6 vs GPT-5.5 vs Gemini 3.1 Pro: 2026 model comparison
Every few weeks, a new model comes out, and every time the same questions arise: should I switch? what changed what to pay for?
Thinking Machines Releases Inkling 2 Small: a 594B MoE Model That Can Fine-Tune Other Models
Thinking Machines released Inkling 2 Small, a 594B-parameter MoE model that leads among open models in several benchmarks and demonstrated autonomous fine-tuning of Qwen3.5-4B.
Four Resume Tips from OpenAI’s Head of Applied AI Development
Wulfie Bain, OpenAI’s head of applied AI development, shared four resume rules: keep it short, remove irrelevant details, highlight outcomes, and lead with the strongest achievements.