The new Laguna S 2.1 model represents a significant advancement in AI capable of performing longer horizon work and utilizing reasoning effectively. This model is a 118B parameter Mixture-of-Experts (MoE) with 8B activated parameters per token, supporting a context window of up to 1 million tokens. It was developed from training to launch in under nine weeks and demonstrates strong performance on long-horizon coding benchmarks, rivaling models many times its size.

Laguna S 2.1 excels in agentic coding tasks, achieving 70.2% on Terminal-Bench 2.1 with its thinking mode enabled. Its compact size makes it suitable for complex work on local machines. The model also shows promise on the Deep SWE benchmark, scoring 40.4% in its agent harness, though this evaluation uses a different harness than the official Deep SWE leaderboard, affecting direct comparability.

Full trajectories for every trial in the final evaluation set are available at trajectories.poolside.ai, offering transparency into the model's behavior. The developers highlight the model's resourcefulness, citing an example where Laguna S 2.1 built a working HTML/CSS rendering engine from scratch and validated its output against a real browser, all within a 50-minute session without human intervention.