Laguna XS 2.1 is Poolside's coding-agent model in the 33B-total, 3B-activated Mixture-of-Experts category, succeeding the April 2026 Laguna XS.2 release and arriving as part of the broader Laguna family. The architecture combines a mixed sliding-window and global attention design, which is what makes it practical to push a context window into the hundreds of thousands of tokens while keeping a relatively small active footprint per token. Because only a few billion parameters fire on any given token, the model is positioned for agentic coding work where many tool calls, long file contexts, and extended reasoning traces are the norm rather than the exception.

For deployment, the open weights are published on Hugging Face and run cleanly through the vLLM Recipes stack, where a BF16 variant fits on a single 80GB GPU and verified hardware spans recent NVIDIA accelerators (H100, H200, B200, GB200, B300, GB300) and AMD MI300X, MI325X, and MI355X cards. The recipe wires up reasoning and tool calling through Poolside-specific vLLM parsers, so agents can stream chain-of-thought output alongside structured function calls without extra glue code. In qualitative fit, that combination of a long context, an MoE efficiency profile, and first-class agent tooling makes Laguna XS 2.1 well suited to repository-scale code generation, multi-step refactors, and other long-horizon software engineering tasks rather than short, single-shot completions.