Laguna XS.2: Our Lightest and Quickest Agentic Coding Model
Laguna XS.2 solves the speed and cost challenges of running advanced AI agents. It is a second-generation model with 33 billion total parameters, but it only activates 3 billion per calculation to keep performance fast. The model was trained in-house on 30 trillion tokens and uses advanced training methods learned from Laguna M.1. Because of this, it works best as the core logic behind fast, adaptive coding agents.
Distributed under an Apache 2.0 license, this open-weight model gives developers complete control. You are free to evaluate, fine-tune, compress or deploy the weights exactly how you want.