AI coding startup Poolside released Laguna S 2.1 on July 21, an open-weight model built for agentic software engineering that the company says outperforms rivals many times its size.
Laguna S 2.1 is a Mixture-of-Experts model with 118 billion total parameters but only 8 billion active per token, letting it run efficient inference while handling a context window of up to 1 million tokens. According to Poolside, the model scored 59.4% on SWE-Bench Pro — within 1.2 percentage points of Moonshot AI’s much larger Kimi K3 — and outperformed DeepSeek-V4-Pro-Max, a 1.6-trillion-parameter model, on several coding benchmarks. On Terminal-Bench 2.1, enabling the model’s “thinking” mode lifted its score from 60.4% to 70.2%.
Poolside credited the gains less to raw scale than to training changes: longer rollout budgets, more verification steps during reinforcement learning, and training across multiple agent harnesses to avoid overfitting to any one of them. “What we’ve done in this model is not necessarily add more intelligence, but improve the behaviors that lead to a more capable model — more verification, less taking things for granted, not declaring victory early, and being more persistent,” said Pengming Wang, Poolside’s co-head of applied research.
As case studies, Poolside said the model built a working HTML/CSS rendering engine from scratch in a single session, cut memory use in the company’s internal testing harness by 71%, and independently reproduced a solution to a 50-year-old open math problem, Erdős problem #397.
The weights are released under Poolside’s OpenMDW-1.1 open license and available on Hugging Face, with a free hosted endpoint on OpenRouter and support in common inference tools including vLLM, Ollama and SGLang. It is the third Laguna-family release in three months, following two smaller models in April and a July 2 update, and took under nine weeks to train.
The release adds to a wave of open-weight coding models this year, a trend Nvidia’s Jensen Huang recently rallied dozens of companies behind, even as rival lab Moonshot AI has had to pause new subscriptions for its own open model, Kimi K3, amid a GPU shortage.