Train a 120B MoE Model on a Single Node

Ringkasan pemeriksaan

Meet LLM, an open-source training pipeline that scales a 2B seed into a 120B Mixture-of-Experts model on a single node. Using TurboQuant Pre-Training, it expands across four stages to 460 routed experts with just six billion active parameters. Stop renting massive clusters. Now, frontier-scale MoE research fits on accessible hardware.