Poolside's Laguna S 2.1 is a small model that beats bigger ones at coding

Poolside released Laguna S 2.1, an open-weight coding model with 118 billion total parameters but only 8 billion active per token, and the interesting claim is efficiency. On Terminal-Bench 2.1, a test of agents working in a real terminal, it scores 70.2%, which Poolside says beats DeepSeek-V4-Pro Max at 64% even though that model carries 1.6 trillion parameters, roughly fourteen times more. It also posts 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro, and handles up to a 1 million token context. The whole model was trained in under nine weeks.

Poolside is direct about what it changed. The team writes that it did not try to add raw intelligence so much as improve behavior: more verification, less taking things for granted, not declaring victory early, and more persistence on long tasks. That shows up in the benchmark math too. Turning on the model's thinking mode lifts its Terminal-Bench score from 60.4% to 70.2%, so a large part of the gain comes from the model checking its own work rather than from scale. As a demonstration, Poolside says the model re-derived a proof to Erdős problem #397, and in a separate run built a working HTML and CSS browser engine in JavaScript in 50 minutes across 181 steps.

The weights are open under the OpenMDW-1.1 license, and hosted access runs about $0.10 per million input tokens and $0.20 per million output on a paid endpoint, with a free 256K-context tier. You can read Poolside's writeup here.

Why it matters

If you build coding agents, a small open model that matches far larger ones on terminal tasks and costs a few cents per million tokens is worth benchmarking on your own repositories, since running it on your own hardware becomes realistic at that size.