Laguna M.1/XS.2 Technical Report
A new technical report on arXiv details the introduction of Laguna M.1 and Laguna XS.2, two Mixture-of-Experts (MoE) foundation models developed for long-horizon, agentic coding tasks. Laguna M.1, the larger model, features 225.8 billion total parameters with 23.4 billion activated per token, while Laguna XS.2 is a more compact model with 33.4 billion total parameters and 3 billion activated per token. Both models were trained from scratch using an internal "Model Factory" system, described as a tightly-integrated stack of versioned data, training, evaluation, and inference components designed to industrialize model development. The report outlines the principles and design choices of this factory, along with the end-to-end training process covering pre-training data, architecture, post-training stages, evaluation, and quantization, emphasizing a systematic approach to model creation. The models demonstrate competitive performance against state-of-the-art open models within their respective weight classes across several agentic software engineering and terminal benchmarks, including SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0. This performance suggests their utility in complex coding environments. Significantly, the weights for Laguna XS.2 have been released under the Apache 2.0 license and are available on Hugging Face, providing developers with an accessible, capable model for integrating into agentic coding applications and experimenting with its capabilities.
Developers gain access to a new, openly licensed Mixture-of-Experts model specifically designed for agentic coding, potentially enhancing automated software development workflows.