Laguna is a family of open Mixture-of-Experts (MoE) foundation models from Poolside built specifically for agentic coding and long-horizon software work. The lineup launched on April 28, 2026 with two models: Laguna XS.2, a 33B-total / 3B-activated model released under the Apache 2.0 license that is compact enough to run locally on a Mac with 36 GB of RAM via Ollama, and Laguna M.1, a 225B-total / 23B-activated model trained from scratch on 30T tokens. On benchmarks, XS.2 scores 68.2% on SWE-bench Verified and M.1 reaches 72.5%, with strong SWE-bench Pro results as well. Poolside also ships 'pool,' a lightweight terminal-based coding agent, and a dual Agent Client Protocol (ACP) client-server — the same environment it uses internally for agent reinforcement-learning training and evaluation.