American AI startup Poolside launches free, high-performing open model Laguna XS.2 for local agentic coding

AI startup Poolside has launched two new Laguna large language models, M.1 and XS.2, alongside new developer tools "pool" and "shimmer." The Laguna M.1 is a proprietary 225-billion parameter MoE model for complex software engineering, while Laguna XS.2 is an Apache 2.0 open-licensed 33-billion parameter MoE model designed for efficient, local agentic coding. Poolside is temporarily offering the larger M.1 for free via API. Both models were trained from scratch using Poolside's "Model Factory," which employs the Titan software, Muon optimizer for 15% faster learning, and AutoMixer for optimal data curation from 30 trillion tokens, including 13% synthetic data. The models undergo Reinforcement Learning in a "virtual gym" to solve real software engineering problems. Benchmarks show Laguna models perform strongly, with M.1 achieving 46.9% on SWE-bench Pro and 72.5% on SWE-bench Verified. Laguna XS.2, despite its smaller size, nearly matches M.1 with 44.5% on SWE-bench Pro, surpassing models like Claude Haiku 4.5. This efficiency is attributed to Poolside's focus on agentic RL and synthetic data. Running XS.2 locally requires specific hardware: 36 GB unified memory for Apple Silicon, or 24-32 GB VRAM for PCs with 4-bit quantization, along with 20-70 GB storage. Poolside recommends using Ollama or their "pool" agent for optimal local execution. The "pool" agent is a terminal-based coding tool designed for developers' local environments, functioning as an Agent Client Protocol (ACP) server. "Shimmer" is a cloud-native, mobile-optimized development environment with an instant-on VM, integrating the Poolside Agent for web app, API, and CLI iteration. A demonstration showed "shimmer" running a full dev environment on a smartphone, highlighting its portability and potential for untethered engineering. The decision to open-license XS.2 aims to support builders and the wider research community, allowing for private, offline deployment in high-security environments, a key advantage for government and public sector clients.