Laguna M.1/XS.2 Technical Report: Mixture-of-Experts Models for Agentic Coding

Forum topic · 小凯 · 2026-05-29

Summary

This forum post summarizes the technical report for Laguna M.1 and Laguna XS.2, two Mixture-of-Experts (MoE) foundation models built for long-horizon, agentic coding. M.1 has 225.8B total parameters (23.4B activated per token), while XS.2 has 33.4B total parameters (3B activated). Both models were trained from scratch end-to-end inside an internal system called the Model Factory: a tightly-integrated, versioned stack of data, training, evaluation, and inference components that industrializes model development. The report details design principles of the Model Factory and the full training pipeline, covering pre-training data and architecture, post-training stages, evaluation, and quantization. On agentic software engineering and terminal benchmarks—including SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0—M.1 and XS.2 are competitive with state-of-the-art open-weight models at their respective sizes. Laguna XS.2 weights have been open-sourced under the Apache 2.0 license on Hugging Face. The paper is available on arXiv (2605.27605).

Overview

This forum post introduces the Laguna M.1/XS.2 Technical Report (arXiv: 2605.27605), authored by Julien Abadji, Marah Abdin, Connor Adams, et al.

Key details

Models: Two Mixture-of-Experts foundation models designed for long-horizon, agentic coding.
Laguna M.1: 225.8B total parameters, 23.4B activated per token.
Laguna XS.2: 33.4B total parameters, 3B activated per token.
Training system: Both models were trained from scratch end-to-end inside the same internal system, referred to as the Model Factory — a tightly-integrated stack of versioned data, training, evaluation, and inference components that turns model development into an industrial process.
Report contents: The paper describes the principles and design choices of the Model Factory and details the end-to-end training process, including pre-training data and architecture, post-training stages, evaluation, and quantization.
Benchmarks: On agentic software engineering and terminal benchmarks (SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0), M.1 and XS.2 are competitive with state-of-the-art open-weight models at their respective weight classes.
Open source: Laguna XS.2 weights are released under the Apache 2.0 license on Hugging Face.

Original abstract (excerpt)

> We present Laguna M.1 and Laguna XS.2, two Mixture-of-Experts foundation models built for long-horizon, agentic coding... On agentic software engineering and terminal benchmarks (SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0) M.1 and XS.2 are competitive with state-of-the-art models...

---

*Auto-collected on 2026-05-29.*