What if you could train a cutting-edge AI model with hundreds of billions of parameters—without shelling out millions on premium GPUs or running a high-performance data center? In this episode, we crack open a groundbreaking report from Ant Group's AI team, revealing how they achieved exactly that.

We take you inside the tech behind Linglight and Ling Plus, two powerful mixture-of-experts (MoE) models that challenge the status quo. With parameter counts rivaling industry giants, these models deliver elite-level performance while operating on modest hardware setups. How? Through a revolutionary cocktail of architectural innovation, smart hardware selection, and next-gen training strategies.

From optimizing model frameworks with DL Rover and the "Eat It" elastic training technique, to scaling across low-cost devices and building robust debugging tools like XPU Timer, this episode unpacks the technical wizardry behind the scenes. We also dive deep into anomaly detection, adaptive tool learning using knowledge graphs, and evaluation tools like Flood—all designed to squeeze out maximum efficiency and stability from limited resources.

But this is more than just a technical deep dive—it's a blueprint for democratizing AI. Whether you’re a researcher, engineer, or just AI-curious, this episode lays out how massive models like Ling Plus (trained on 9 trillion tokens, no less) can be built without Wall Street-level budgets.

If you've ever felt locked out of high-end AI development, this episode is your key. Tune in and discover how the next era of large language models is being built—not just bigger, but smarter, faster, and more accessible.

Read more: https://arxiv.org/pdf/2503.05139