Poolside Releases Open Source 118 Billion Parameter Code Model, Self-Tested Against 1.6 Trillion Parameter DeepSeek

According to the DynaBeat Monitor, Poolside, a dedicated programming Agent, has open-sourced the MoE model Laguna S 2.1. The model has a total of 118 billion parameters, with only 80 billion activated per instance, supporting context for 1 million tokens. It can be run on a single NVIDIA DGX Spark after quantization.

During Poolside's self-testing, Laguna S 2.1 achieved a score of 70.2% on Terminal-Bench 2.1, surpassing the 64.0% score of DeepSeek-V4-Pro-Max with a total of 16 trillion parameters and 490 billion activated parameters. It also outperformed SWE-Bench Multilingual, SWE-Bench Pro, and DeepSWE.

However, there is still a significant gap compared to Kimi K3. Kimi K3, with a total of 28 trillion parameters and 500 billion activated parameters, scored 88.3% on Terminal-Bench 2.1 and 69% on DeepSWE. In contrast, Laguna S 2.1 scored 70.2% and 40.4%, respectively.

These scores were not obtained under the same testing environment. Poolside used its proprietary Agent framework, while competitors achieved their publicly reported top scores. A key feature of Laguna S 2.1 is its ability to achieve long-range programming capabilities on a local single machine using fewer activated parameters.