Poolside has launched Laguna S 2.1, an open-weight coding model presented as the Western counterpart to DeepSeek and Qwen.

      Poolside has introduced Laguna S 2.1, a 118-billion-parameter open-weight coding model designed for agentic coding. The San Francisco startup asserts that this model competes with or surpasses larger models. Utilizing a mixture-of-experts architecture with eight billion active parameters per token, it is compact enough to operate on a single Nvidia DGX Spark desktop system, and the model weights can be accessed on Hugging Face under the Linux Foundation’s OpenMDW license.

      In evaluations such as Terminal-Bench and SWE-Bench Pro, Laguna S 2.1 achieved scores of slightly over 70 percent and almost 60 percent, respectively, comparable to or better than models from DeepSeek, Nvidia, and Thinking Machines that have two to eight times the number of active parameters. Poolside admits that the model is “not yet at the frontier,” as closed-source systems from OpenAI and Anthropic perform significantly better on the same benchmarks.

      This release is positioned as a response to the leadership of Chinese labs in the open-weight sector, where models like those from DeepSeek, Alibaba’s Qwen family, and Moonshot’s Kimi have led for over a year. Prior to this launch, no Western lab had released an open-weight model within the 118-billion-parameter range for 11 months, as indicated by the company. Forbes reported that Poolside aimed to provide Western businesses and governments with a self-hosted alternative that prevents data from being sent to foreign entities.

      Founded in 2023 by Jason Warner, the former chief technology officer at GitHub, and Eiso Kant, Poolside raised $500 million in a Series B round in October 2024 at a valuation of $3 billion, with investments from Nvidia and eBay. A planned $2 billion Series C, which would have valued the company at $14 billion, fell through in April 2026 after CoreWeave withdrew from a joint data center project in Texas. The company now caters to government, defense, and other highly regulated organizations through its API and agent harness.

      Poolside claims it developed the model using its internal Model Factory platform, which automates architecture search and reinforcement learning from code execution, completing training in under four weeks on 4,000 Nvidia H200 GPUs. The smaller model, Laguna XS, was launched three weeks prior, and the company indicates it releases new models approximately every five weeks. To showcase long-horizon reasoning, Poolside published a demonstration of the model successfully solving a combinatorics problem that previously only the most advanced models could resolve.

      The strategy is to convince enterprises to run a robust coding model on their hardware instead of relying on a closed API, which hinges on Laguna S 2.1's performance in real-world applications aligning with its benchmark results. Poolside’s own findings show the model trailing closed-source leaders by about 10 to 15 percentage points on Terminal-Bench, a difference that is significant for clients weighing the benefits of self-hosting.

      The company's ability to close this performance gap in the next iteration, while contending with rapidly advancing Chinese open-weight models, will determine whether the Western open-weight disparity is a temporary issue or a more persistent problem.