Inside Poolside AI's Infrastructure for Training the 118B Laguna S Model

Eiso Kant, co-CEO of Poolside AI, detailed how a small engineering team built a "model factory" to train Laguna S, a 118-billion parameter Mixture of Experts (MoE) model. This model reportedly outperforms Thinking Machines' ~1-trillion parameter open-weights model, Inkling. It demonstrates that highly optimized infrastructure and training methodologies can allow smaller, specialized models to outperform much larger models, lowering the barrier to entry for training state-of-the-art AI. This is particularly significant for open-weight coding models where efficiency and cost-effectiveness are critical. The Laguna S model utilizes a Mixture of Experts (MoE) architecture, allowing it to achieve high performance while remaining computationally efficient. The engineering team focused on building a scalable "model factory" pipeline to streamline training processes rather than just scaling up raw compute.