Laguna XS.2 and M.1: A Deep Dive

The first two models in the Laguna family, M.1 and XS.2, have been released. M.1 is a foundation model that completed pretraining at the end of last year, while XS.2 is a smaller, second-generation model and the first weights-open release. Based on the judgment that agents that write and execute code are a more expressive interface than fixed tool calling, they presented the models together with the ACP runtime that runs them.

• Laguna M.1: A MoE model with 225B total / 23B activated parameters, trained entirely in-house on 30T tokens. It used 6,144 NVIDIA Hopper GPUs, and scored 46.9% on SWE-bench Pro and 40.7% on Terminal-Bench 2.0.
• Laguna XS.2: A second-generation MoE with 33B total / 3B activated parameters. It scored 44.5% on SWE-bench Pro and 30.1% on Terminal-Bench 2.0, and its weights are released under Apache 2.0.

Both models are available for free for a limited time via the API and OpenRouter, and XS.2 is also supported on Ollama and NVIDIA TensorRT-LLM. An NVFP4 version is also provided, which is expected to perform strongly on the Blackwell architecture. The agent harness released alongside them is an Agent Client Protocol (ACP) server, the same execution layer used for agent RL training and evaluation.

The entire training stack was built in-house. On top of the Titan codebase and their own agent RL infrastructure, web data was optimized for both quality and diversity — rather than keeping only high-quality data, some mid- and lower-quality buckets were also retained to reduce bias toward STEM and reasoning. Synthetic data made up about 13% of the final mix, with 4.4T+ synthetic tokens used overall. They also reconfirmed FineWeb's hypothesis that global deduplication excessively removes high-quality data.

For data mix optimization, they used AutoMixer, which trains a surrogate regressor by running about 60 proxy models. Their in-house distributed implementation of the Muon optimizer reached the same loss with about 15% fewer steps than AdamW, and Newton-Schulz-based orthogonalization was handled through communication and distribution to reduce bottlenecks. Evaluation was conducted using Laude Institute's Harbor Framework and their own agent harness, mostly under sandbox conditions of up to 500 steps and 8GB RAM/2 CPUs. Only Terminal-Bench 2.0 was measured with 48GB RAM/32 CPUs.