Laguna S 2.1: 118B MoE model with 1M-context window and strong agentic performance
Summary
Poolside announces Laguna S 2.1, a 118B Mixture-of-Experts model with a 1M-token context window, designed for long-horizon reasoning and agentic coding tasks. The post presents benchmark results across multiple suites (Terminal-Bench 2.1, SWE benchmarks, DeepSWE, etc.), multiple case studies, training/post-training details, and deployment options via open sources and hosted platforms. It also discusses evaluation methodology, limitations, and the model’s emphasis on persistence, verification, and tool use.