Laguna M.1: Our Most Capable Model for Agentic Coding
Laguna M.1 solves the complication of running massive, high-performing AI coding workflows efficiently. This model holds 225 billion total parameters but only activates 23 billion per calculation. It was trained from scratch on 30 trillion tokens using a massive cluster of 6,144 NVIDIA H200 GPUs.
Laguna M.1 delivers its best performance when powering advanced coding agents. The model was built specifically for agentic tasks by using custom data, a dedicated training codebase and specialized reinforcement learning.