Poolside's 8B model beats trillion-parameter giants

poolside released laguna s 2.1, an open-weight coding model that leverages 8 billion active parameters via a mixture-of-experts architecture to outperform competitors 10 to 20 times its size on benchmarks like terminal-bench 2.1 and deepswe. the model demonstrates that verification and persistence—such as not declaring victory early—can substitute for raw parameter scale, achieving pass rates that surpass trillion-parameter models. in tests, it successfully built a browser engine from scratch and independently solved erdős problem #397, a math problem open since 1975. this release signals a shift toward efficiency and agentic behavior in coding models, though a larger laguna model remains in pre-training.