Laguna by Poolside is the open-weight foundation model family from Poolside, a company whose stated goal is to build software agents on the path to artificial general intelligence. The Laguna line currently spans two models: Laguna s 2.1, a 118B-parameter model with 8B active parameters and a 1M token context window that Poolside positions as frontier-class reasoning at mid-size cost, and Laguna xs 2.1, a 33B-parameter model with 3B active parameters and a 256K context window designed to be small enough to run on-device. Both are tuned for agentic coding rather than generic chat.,What makes the family notable is that the weights are published openly — you can download them on Hugging Face — while Poolside continues to develop the runtime it uses to train and operate agents through its own product experiences, currently in research preview. Poolside also works directly with government and enterprise partners, offering an on-premise and secure deployment posture for organisations that need to run frontier AI inside their own environments rather than through a third-party API.,Poolside pairs its model releases with unusually detailed public research. Its Model Factory blog series documents how the company trains, evaluates and iterates on foundation models, covering topics from GPU-to-GPU weight transfers and automated architecture ablations to reinforcement learning from code execution at scale. Other posts examine benchmark hacking and reward hacks, and hardware optimisations such as C2C activation offloading on Grace Blackwell superchips. For developer teams evaluating agentic coding models, Laguna is interesting primarily as an open, inspectable alternative with a serious research organisation behind it.