Poolside Releases Laguna S 2.1, Open-Weight Model Specialized for Coding

Poolside, an AI startup based in San Francisco, has released a new AI model for coding called Laguna S 2.1. The model weights have been immediately released on Hugging Face under the OpenMDW-1.1 license, allowing anyone to freely download and use them. After three years of primarily providing models to government and defense agencies, the company has released its first general-purpose open-weight model to the public.

Recently, adoption of "open-weight models" has been accelerating among enterprises and developers. This is because they can run on companies' own infrastructure without relying on cloud services, offering advantages in security and cost. However, the leading choices of open models suitable for such use have been centered on models from China, such as DeepSeek, Qwen, and Kimi. According to Poolside's announcement materials, this marks the first time a Western research institution has deployed an open-weight model in the same scale tier since OpenAI's gpt-oss-120b in August 2024, a gap of 11 months.

Laguna S 2.1 employs an architecture called MoE (Mixture of Experts). This approach maintains 118 billion total parameters for the model, but only activates 8 billion parameters during actual processing, thereby achieving large-scale performance with minimal computational cost. It also features a context window capable of processing up to 1 million tokens (equivalent to approximately 500,000 to 750,000 characters in Japanese) at once.

In terms of performance, the model achieved a score of 70.2% on Terminal-Bench 2.1, a benchmark for evaluating extended terminal operation tasks, surpassing DeepSeek-V4-Pro-Max with 1.6 trillion parameters (64.0%) and Nvidia's Nemotron 3 Ultra with 550 billion parameters (56.4%), according to the company's announcement. On SWE-Bench Multilingual, a widely used coding evaluation standard, it scored 78.5%, and on the SWE-Bench Pro public dataset, it scored 59.4%. The fact that the model exceeded much larger competitors in scores represents a result that questions the relationship between scale and performance.

Development speed is also noteworthy. Pre-training began on May 22, with release completed in less than 9 weeks. The GPU used was Nvidia's H200, with 4,096 units. Considering that mainstream AI model development typically requires several months to over a year, the company has released three models consecutively within these three months.

Poolside co-CEO Jason Warner stated in the announcement, "The West needs open-weight models that are trustworthy, that companies can run themselves, and that they can build on top of." Additionally, co-founder and fellow co-CEO Iso Kanto posted on X (formerly Twitter) that "Intelligence (AI capability) should become a commodity, and it will." This reflects the view that for open models to survive, they must achieve performance equal to or exceeding closed models.

This development demonstrates that the competition for AI dominance transcends mere performance rivalry and intertwines with the geopolitical question of "which country or bloc can supply trustworthy open AI." Poolside's core business centers on enterprise services that deploy models in secure environments. The public release as an open-weight model can be understood as having strategic significance in expanding options for Western enterprises and government agencies to leverage AI while avoiding dependence on Chinese models.