AI startup Sand.ai has open-sourced its core audio-video era expertise stack over three consecutive days on GitHub.
The releases embody: • daVinci-MagiHuman, a 15B-parameter multimodal era mannequin • MagiAttention v1.1.0, a distributed consideration module • MagiCompiler, a unified training-and-inference compilation framework
Sand.ai was based by former Microsoft Analysis Asia scientist Cao Yue, with workforce members beforehand contributing to the event of Swin Transformer.
The corporate focuses on autoregressive world fashions, and has beforehand launched fashions resembling Magi-1 (video era) and GAGA-1 (audio-visual era).
The open-source initiative goals to share developments throughout mannequin structure, compute infrastructure, and compilation frameworks, contributing to foundational infrastructure for video era.
Supply: Minds in AI
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the most recent breakthroughs, get unique updates, and join with a worldwide community of future-focused thinkers.
Unlock tomorrow’s traits at present: learn extra, subscribe to our e-newsletter, and change into a part of the NextTech neighborhood at NextTech-news.com
