Laguna XS.2

Laguna XS.2 is Poolside’s local-oriented sparse Mixture-of-Experts model for agentic coding and long-horizon work. It has 33 billion total parameters, with only 3 billion activated per token. Sliding-window and global attention are mixed in a three-to-one layout, while FP8 KV caching reduces memory per token. Native reasoning supports preserved thought between tool calls and can be enabled or disabled per request.

Poolside describes the checkpoint as runnable on a Mac with 36 GB of RAM and documents support for vLLM, Transformers, SGLang, and Ollama. Laguna M.1 is the much larger family member. The context window holds 262,144 tokens.