Poolside open-sources the 118B-parameter code model Laguna S 2.1, outperforming DeepSeek in self-tests.

According to Beating Monitor, Poolside, specializing in programming agents, has open-sourced its MoE code model, Laguna S 2.1. The model has a total of 118 billion parameters, activating only 8 billion per inference, and supports a context length of up to 1 million tokens. After quantization, it can run on a single NVIDIA DGX Spark. In Poolside’s internal tests, Laguna S 2.1 achieved 70.2% on Terminal-Bench 2.1, outperforming DeepSeek-V4-Pro-Max, which has 1.6 trillion total parameters and 49 billion activated parameters, at 64.0%. Laguna S 2.1 also scored higher on SWE-Bench Multilingual, SWE-Bench Pro, and DeepSWE. However, it still lags significantly behind Kimi K3, which has 2.8 trillion total parameters and 500 billion activated parameters, achieving 88.3% on Terminal-Bench 2.1 and 69% on DeepSWE, compared to Laguna S 2.1’s 70.2% and 40.4%, respectively. These scores were not retested under identical conditions; Poolside used its own agent framework, while competitors’ results reflect publicly reported top performances. The key advantage of Laguna S 2.1 is achieving strong long-context programming capabilities with fewer activated parameters, enabling local execution on a single machine.