[Poolside Open-Sources 118 Billion Parameter Code Model, Outperforming 1.6 Trillion Parameter DeepSeek in Self-Test] According to monitoring by Dongcha Beating, 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, with only 8 billion activated per run, supporting a context of up to 1 million tokens. After quantization, it can run on a single NVIDIA DGX Spark machine. In Poolside's self-tests, Laguna S 2.1 scored 70.2% on Terminal-Bench 2.1, surpassing the 64.0% achieved by DeepSeek-V4-Pro-Max, which has a total of 1.6 trillion parameters and 49 billion activated parameters. It also outperformed in SWE-Bench Multilingual, SWE-Bench Pro, and DeepSWE benchmarks. However, it still lags significantly behind Kimi K3. Kimi K3 has a total of 2.8 trillion parameters and 50 billion activated parameters, scoring 88.3% on Terminal-Bench 2.1 and 69% on DeepSWE. In comparison, Laguna S 2.1 scored 70.2% and 40.4%, respectively. These scores were not retested under the same environment. Poolside uses its proprietary agent framework, while competitors' scores reflect their publicly reported highest results. The standout feature of Laguna S 2.1 is its ability to achieve long-range programming capabilities with fewer activated parameters, making it runnable on a single local machine. [Original Link]