Laguna XS 2.1 is poolside's open-weight coding agent model built around a Mixture-of-Experts design, pairing 33 billion total parameters with only about 3 billion active per token so the heavy lifting is gated to specialized experts while keeping inference lean. That ratio is what allows the model to be quantized down to roughly 16–20 GB of VRAM at INT4, putting real local agentic coding within reach of a workstation rather than a data center. The release is positioned as a direct successor to the earlier Laguna XS.2, refining the family rather than reinventing it, and it continues poolside's broader push into open models paired with a runtime that supports training and serving agents.
On practical workloads, the upgrade over XS.2 shows up most clearly in software engineering evaluations and terminal-style tasks, where poolside reports a 5.4 point gain on SWE-bench Multilingual alongside stronger terminal benchmarks and an expanded 262K context window that comfortably handles long agent traces and multi-file refactors. The model is aimed squarely at agentic coding and long-horizon work on a local machine, which makes it a natural fit for developers who want a self-hosted reasoning and tool-using partner without depending on a hosted frontier API. Its combination of open weights, MoE efficiency, and benchmark gains over the previous Laguna release makes it one of the more compelling options in the open-weight coding agent space.
