Laguna XS 2.1 is Poolside's compact coding-agent model built around a Mixture-of-Experts design with 33B total parameters and 3B activated per token, positioning it as a smaller sibling within the Laguna family that also includes the larger 118B-total / 8B-active Laguna S 2.1. The model is published openly on Hugging Face under the Poolside organisation and is redistributed as a third-party offering through the NVIDIA NIM catalogue, with global deployment availability and governance under the OpenMDW License. Distribution through NIM alongside an open Hugging Face release means developers can either run the weights locally or pull the model from an enterprise-friendly inference endpoint, giving flexibility for both prototyping and production-grade agentic coding pipelines.

The intended use case for Laguna XS 2.1 centres on software-engineering workloads: agentic coding, terminal-style tasks, tool-use workflows, and long-horizon programming sessions on local machines. Compared with the heavier Laguna S 2.1, which records 70.2% on the Terminal-Bench 2.1 agentic coding benchmark, Laguna XS 2.1 trades raw benchmark scale for a leaner active-parameter footprint that fits comfortably on developer hardware while still inheriting the family's long-context behaviour and code-specialised training focus. Practical fit therefore maps well to individual developers and small teams who want a locally runnable coding agent tuned for multi-step repository work, rather than to scenarios that demand the largest possible active-parameter count.