Laguna XS 2.1

Name: Laguna XS 2.1
Author: Poolside

Poolside · 2026-07-02 · 33.4B parameters

Laguna XS 2.1 is Poolside's open-weight Mixture-of-Experts model for agentic coding and long-horizon work, described by Poolside as 33B total parameters with 3B activated per token. It upgrades Laguna XS.2, supports per-request enable/disable of interleaved thinking, and uses 10 global-attention and 30 512-token sliding-window-attention layers across an exact 262,144-token context.

Benchmark scores

Benchmark	Score
AHK-Eval	25.0
AHK-Eval Algorithms	16.7
AHK-Eval Data Structures	16.7
AHK-Eval Date and Time	16.7
AHK-Eval Easy	41.7
AHK-Eval Hard	25.0
AHK-Eval Hidden Cases	26.0
AHK-Eval Mid	16.7
AHK-Eval Numbers	50.0
AHK-Eval Parse Success	72.2
AHK-Eval Regex	66.7
AHK-Eval Strings	0.0
BuseyBench SVG	2.0
Eval Connections	30.0
Eval Connections — Guess Accuracy	36.3
Eval Connections One-Shot Group Accuracy	21.3
Eval Connections One-Shot Puzzle Solves	10.0
Eval Connections One-Shot Trap Bonus	5.0
Eval Connections One-Shot with Traps	17.0
GBENCH One-shot Coding GScore	21.8