Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual

Poolside has released an open-weight model designed for coding tasks that uses a sparse architecture, meaning only a fraction of its total parameters activate per step, allowing it to perform comparably to much larger models while remaining cheaper to run. The model scores competitively on coding benchmarks against larger closed and open models, ranking first among open models with disclosed sizes on some benchmarks. It can run on a single GPU at lower precision levels and was trained in under nine weeks. The model includes a "thinking mode" that enables it to spend more computational effort on complex problems, though this increases token usage substantially.