Poolside launches Laguna S 2.1, an efficient 118B-parameter coding model

Overview of Laguna S 2.1 Launch

Poolside launched its Laguna S 2.1 model, touted as its most capable release thus far. The model employs a Mixture-of-Experts system that selectively activates a smaller subset of its parameters, allowing it to optimize efficiency compared to previous large-scale models. This approach aims to position the smaller lab as a serious competitor in the AI space dominated by larger organizations.

Performance Metrics of Laguna S 2.1

Laguna S 2.1 has achieved a score of 70.2% on Terminal-Bench 2.1, significantly surpassing larger models such as the 1.6-trillion-parameter DeepSeek-V4-Pro-Max, which scored 64.0%. Other benchmarks include 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro's public dataset, indicating strong performance in context-heavy coding tasks.

Significance of Open-Weight AI

The release comes at a critical time as the conversation around open-weight AI models intensifies, particularly concerning the options available from Western labs. Poolside emphasizes the need for reliable open-weight models, as rival models from Chinese labs dominate the current landscape. The launch of Laguna S 2.1 aims to fill a reported gap in the market that has seen a lack of new open-weight models from Western developers.

Rapid Development Cycle

The development timeline for Laguna S 2.1 was notably swift, taking less than nine weeks from the beginning of pre-training on May 22 to its public release. This speed contrasts with typical industry timelines requiring months or years to iterate on flagship models, suggesting Poolside's commitment to rapid innovation in the AI sector.