Laguna XS 2.1 sits in Poolside's "Laguna" line of coding agent models and is presented as a step forward from Laguna XS.2, which shipped in April 2026. It belongs to the 33B-A3B category, a Mixture-of-Experts configuration with around 33 billion total parameters but roughly 3 billion activated per token, which the company frames as a compact footprint aimed at agentic coding on local or single-host setups. Training lineage is anchored in Poolside's agent-oriented pipeline, including data automixing and asynchronous off-policy agent reinforcement learning described in the accompanying technical report, giving the model a software-engineering focus rather than a general-purpose conversational one.

In qualitative terms, Laguna XS 2.1 is built for software engineering and agentic coding tasks such as multi-step repository work, terminal-style operations, and tool-driven debugging. It supports tool calling and reasoning together, and OpenRouter reports a measurable gain over Laguna XS.2 on SWE-bench Multilingual plus stronger terminal-task performance, which suggests steady progress on long-horizon coding benchmarks rather than a single-axis leap. The open-weights release is distributed on Hugging Face under the poolside organization as an FP8 quantization for fast, cost-efficient inference, shipped under the OpenMDW-1.1 license and aligned with Poolside's Acceptable Use Policy, making it a practical fit for teams that want a self-hostable, code-first model with a 256K context window and up to 32K output tokens for sustained agent sessions.