For the past two years, the best AI models available for self-hosting primarily came from China, with models like GLM, Deepseek, and Quen dominating the open-source landscape. Western alternatives were scarce and often required cloud access or API usage, limiting control and trust for industries like banking, defense, and healthcare. This changed with the release of Laguna S 2.1 by the startup Poolside, a model with 118 billion parameters that activates only 8 billion per word generated. This design allows it to run efficiently on a single desktop without the need for expensive data centers or ongoing API fees, marking a significant milestone for Western open-source AI.
Laguna S 2.1 achieves its performance through a novel approach where a small router selectively activates only the necessary “experts” within the model for each word, keeping the majority inactive. This means it carries the knowledge of a massive model but operates with the computational cost of a much smaller one. The model can process a million tokens in one go, enabling it to understand and work with entire large codebases simultaneously. This efficiency allows Laguna to compete with much larger models on complex coding benchmarks, outperforming some models with over a trillion parameters in tasks that reward understanding and reasoning.
In benchmark comparisons, Laguna S 2.1 excels in multilingual bug fixing and code-based question answering, outperforming GLM 5.2 in some areas despite being significantly smaller. However, GLM still holds a slight edge in other key benchmarks, making the competition a split decision rather than a clear victory. Laguna’s ability to run locally on affordable hardware like an Nvidia DGX Spark or even a standard desktop with sufficient RAM and a mid-range GPU makes it highly accessible. This local capability eliminates ongoing costs and privacy concerns associated with cloud-based models, making it particularly attractive for continuous, heavy use.
Poolside, the company behind Laguna, was founded in 2023 by industry veterans including Jason Warner, former CTO of GitHub. With substantial funding from Nvidia and others, they trained Laguna in under nine weeks using advanced techniques like reinforcement learning in 8-bit precision, which is considered risky but cost-effective. Despite the impressive claims, it’s important to note that all benchmark results currently come from Poolside itself and have yet to be independently verified. Additionally, Laguna still trails behind some closed-source models like Anthropic’s Claude in certain tasks, highlighting that open models have made significant progress but have not yet fully closed the gap.
In summary, Laguna S 2.1 represents a major step forward for Western open-source AI models, offering a powerful, locally runnable alternative to the previously dominant Chinese models. Its innovative architecture allows it to deliver high performance on standard hardware while maintaining full user control and privacy. While it doesn’t surpass all competitors in every benchmark, it provides a practical and trustworthy option for organizations needing robust AI without reliance on cloud services. This development signals a renewed and competitive presence of Western models in the open AI race, making Laguna S 2.1 arguably the best local model available today.
