118 billion parameters, eight billion active, and a local machine

Poolside's Laguna S 2.1 model reduces activation to eight billion parameters per token. An architecture designed for heavy agentic code.

With Laguna S 2.1, Poolside centers a 118-billion-parameter model around a Mixture-of-Experts architecture that activates only eight billion parameters per token. This reduction is intended to allow it to run on a single NVIDIA DGX Spark, while maintaining a context window of up to one million tokens.

The model primarily targets agentic code and long-horizon tasks, with a particular focus on persistence, verification, and the ability to backtrack from an incomplete approach. Poolside claims 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1 using its own evaluation harness. The complete trajectories of the final trials are published to allow for review.

The maximum reasoning mode significantly improves results, but can also lead to unnecessarily long sequences. The company also reports difficulties with certain tool-calling formats used by third-party harnesses and with nested JSON arguments.

Laguna S 2.1 is distributed as open weights under the OpenMDW-1.1 license, with several compression levels available on Hugging Face. It is also accessible via OpenRouter, Poolside's API, and a web interface with no login required.