Laguna XS.2 and M.1 (poolside.ai)

Laguna has unveiled its first two models, Laguna M.1 and Laguna XS.2, as part of the Laguna family, along with their runtime for training and operating AI agents. Laguna M.1 is a powerful 225 billion parameter model designed for high-security applications, achieving impressive benchmarks with 46.9% on SWE-bench Pro. Meanwhile, Laguna XS.2, a smaller model with 33 billion parameters, is notable for being the first open-weight model in the series and is available under an Apache 2.0 license. Users can access both models for free during the preview period, aiming to foster innovation within the AI community. The introduction of these models is significant as they focus on long-horizon tasks and coding capabilities, marking a shift from traditional agent interactions, which rely heavily on fixed toolsets. The Laguna models demonstrate enhanced flexibility by allowing agents to write and execute code, opening the door for more complex and autonomous AI systems. Developed on NVIDIA hardware, the models utilize a sophisticated data curation and training process, incorporating over 30 trillion tokens and advanced techniques like synthetic data augmentation. The release of these models is seen as a critical contribution to the open-weight ecosystem, encouraging developers and researchers to engage with, enhance, and build upon these foundational tools.