en.Wedoany.com Reported - US startup Poolside has released Laguna S 2.1, an open-weight coding model with 118 billion parameters, employing a mixture-of-experts architecture that activates 8 billion parameters per token. Designed specifically for agentic coding, the model is compact enough to run on a single Nvidia DGX Spark desktop system, and its weights have been published on Hugging Face under the Linux Foundation's OpenMDW license.
In two agentic coding evaluations, Terminal-Bench and SWE-Bench Pro, Laguna S 2.1 scored slightly above 70% and close to 60%, respectively. Poolside stated that the model's performance matches or surpasses models from companies such as DeepSeek, Nvidia, and Thinking Machines that have 2 to 8 times more active parameters, but acknowledged that the model has not yet reached frontier levels, with closed-source systems from OpenAI and Anthropic still clearly leading on the same benchmarks.
This release is seen as a direct response to the dominance of Chinese labs in the open-weight space. Over the past year and more, DeepSeek, Alibaba's Qwen series, and Moonshot AI's Kimi have maintained a leading position in this field. According to the company, no Western lab had released an open-weight model at the 118-billion-parameter level in the 11 months prior to this launch. As reported by Forbes, Poolside positions this release as providing Western enterprises and governments with a deployable alternative that allows them to avoid sending data to overseas vendors.
Poolside was founded in 2023 by former GitHub CTO Jason Warner and Eiso Kant. In October 2024, it completed a $500 million Series B funding round at a $3 billion valuation, with support from Nvidia and eBay. In April 2026, a planned $2 billion Series C round at a $14 billion valuation fell through after CoreWeave withdrew from a joint data center project in Texas. The company currently serves government, defense, and other highly regulated organizations through its API and agentic tools.
The company stated that it built the model using its internal Model Factory platform, which automates architecture search and code execution-based reinforcement learning, and completed training in less than four weeks using 4,000 Nvidia H200 GPUs. A smaller model, Laguna XS, was also released three weeks ago, with the company indicating a roughly five-week cadence for new model releases. As a demonstration of long-range reasoning, Poolside released a trace of the model independently solving a combinatorics problem, which until recently could only be solved by the largest frontier models.
Poolside's business model is based on the premise that enterprises will pay to run powerful coding models on their own hardware rather than sending prompts to closed APIs. This expectation depends on whether Laguna S 2.1's performance in production environments matches its benchmark results. The company's own results show the model trailing closed-source leaders by approximately 10 to 15 percentage points on Terminal-Bench. Whether the company can close this gap in the next cycle while competing with rapidly advancing Chinese open-weight models will determine whether the West's gap in the open-weight space is a temporary phenomenon or a structural issue.
