Laguna XS.2
Poolside · 2026-04-28 · 33.4B parameters
Laguna XS.2 is Poolside's open-weight Mixture-of-Experts model for agentic coding and long-horizon work; the released checkpoint contains exactly 33,442,617,088 parameters, while Poolside reports a rounded 3B activated per token. Its 40 layers mix sliding-window and global GQA in a 3:1 ratio, with 48 query heads on global layers and 64 on sliding-window layers, and it supports enabling or disabling interleaved reasoning per request. Poolside trained the model from scratch through pre-training, post-training, and RL; the report says pre-training exceeded 30T tokens and explains that its 256K context extends 128K long-context training by doubling the global-layer RoPE scale without further training.