The American-made open-type camera 'Laguna S 2.1' has arrived, boasting 118 billion parameters, surpassing the 1.6 trillion parameters of the DeepSeek-V4-Pro-Max.

Poolside, an AI development company based in the United States, announced its own model, ' Laguna S 2.1 ,' on July 21, 2026, local time.

Introducing Laguna S 2.1 — Poolside

Today we're releasing Laguna S 2.1, our most capable model to date.

It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.

Capable enough to hold its own against models many… pic.twitter.com/hJr4yQ6VzA

— Poolside (@poolsideai) July 21, 2026

Laguna S 2.1 is a MoE model with a total of 118 billion parameters and 8 billion active parameters, and its context window has a maximum of 1,048,576 tokens. It features both thinking and non-thinking modes and is touted as achieving performance comparable to much larger models. Furthermore, the period from the start of training to release is less than 9 weeks.

The following are the results of Terminal-Bench 2.1, which measures the performance of AI agents. Laguna S 2.1 surpassed DeepSeek-V4-Pro-Max, which has a total of 1.6 trillion parameters, achieving the highest performance among American-made open models and ranking among the top even when compared to Chinese-made open models.

The graph below shows the number of parameters on the horizontal axis and the Terminal-Bench 2.1 score on the vertical axis. It can be seen that Laguna S 2.1 performs better than models of similar size and is a smaller, more efficient model than models with similar performance.

It demonstrates high performance not only in benchmarks but also in real-world tasks. In a task to develop a browser engine from scratch, it successfully rendered HTML and CSS in 50 minutes.

It excels at lengthy tasks that can last several hours. The following shows the average number of output tokens per trial and the task completion rate, demonstrating that increasing the number of output tokens by switching to thinking mode also improves the task completion rate.

Laguna S 2.1 is available as an open model at the following link. The license is OpenMDW-1.1 .

poolside/Laguna-S-2.1 · Hugging Face
https://huggingface.co/poolside/Laguna-S-2.1

Laguna S 2.1 has a size of 236GB and requires multiple GPUs and a large amount of memory to run. Poolside has also released various quantized versions of Laguna S 2.1, and the NVFP4 version can be run locally with DGX Spark or Mac Studio.

Laguna S 2.1 - a poolside Collection
https://huggingface.co/collections/poolside/laguna-s-21