33B total / 3B active (MoE)
128K tokens
Apache 2.0
Poolside's first open-weight model and second-generation XS-class agentic coding release, trained from scratch as a Mixture-of-Experts model for long-horizon software work. Official release materials describe Laguna XS.2 as a 33B-parameter model with 3B activated parameters trained on more than 30T tokens, incorporating lessons from Laguna M.1 across data, synthetic data generation, and agent reinforcement learning. Poolside reports 44.5% on SWE-bench Pro, 68.2% on SWE-bench Verified, 62.4% on SWE-bench Multilingual, and 30.1% on Terminal-Bench 2.0, positioning it as a strong open-weight model in its size class for coding agents. The blog also highlights a training stack built around Poolside's internal Titan codebase, distributed Muon optimizer, asynchronous long-horizon agent RL, and substantial synthetic-data use, with synthetic data contributing about 13% of the final XS.2 training mix and the Laguna family using roughly 4.4T+ synthetic tokens overall. Release materials further note day-one TensorRT-LLM support, an NVFP4 version for NVIDIA Blackwell, and compatibility with Poolside's released agent harness runtime.