← Retour au blog
tech 22 July 2026

Laguna S 2.1: A Leap Forward for MoE Models

Explore Laguna S 2.1, Poolside's groundbreaking model redefining Mixture-of-Experts horizons with unmatched reasoning efficiency.

Article inspired by the original source
Laguna S 2.1 ↗ poolside.ai

Introduction

In the ever-evolving world of artificial intelligence, Laguna S 2.1 marks a decisive turning point. Developed by the Poolside team, this model is a significant step towards the future of Mixture-of-Experts (MoE) models. With 118 billion total parameters, including 8 billion activated per token, Laguna S 2.1 stands out with its ability to handle context windows of up to one million tokens. Here's a look at what makes this model exceptional.

A Model Punching Above Its Weight

Laguna S 2.1 was designed to compete with much larger models, and it does so brilliantly. On long-horizon coding benchmarks, it holds its own against giants like Tencent Hy3 (295B-A21B) and Nemotron 3 Ultra (550B-A55B). For instance, on the Terminal-Bench 2.1, Laguna scored 70.2, outperforming Nemotron and approaching Kimi K3 (88.3). This performance is remarkable given its relatively modest size.

DeepSWEE: An In-Depth Evaluation

One of the distinguishing features of Laguna S 2.1 is its DeepSWEE evaluation system. This system tracks every trial and provides full trajectories for every final evaluation set. This ensures complete transparency and allows users to understand how the model achieves its results. This approach is unprecedented and offers a new level of understanding of AI model performances.

What's Changed in Laguna S 2.1

The development of Laguna S 2.1 was swift, taking less than nine weeks from the start of training to launch. This model benefited from a strong base followed by a series of carefully distributed post-training tasks. The training loop was optimized to maximize efficiency, allowing for a faster process without compromising quality.

Poolside's Bets

With Laguna S 2.1, Poolside has made bold bets. The company has launched three models in three months, each addressing specific market needs. This speed of execution is a testament to Poolside's commitment to constant innovation. Laguna S 2.1 is designed to excel in complex reasoning tasks, paving the way for applications that require deep thinking and long-term data processing.

Conclusion

Laguna S 2.1 is much more than just an update to an existing model. It is a significant advancement that pushes the boundaries of MoE model capabilities. For tech decision-makers and entrepreneurs, this model offers unprecedented opportunities for innovation and process optimization.

Let's discuss your project in 15 minutes.

Laguna S 2.1 Mixture-of-Experts AI models DeepSWEE Poolside
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call