Laguna S 2.1 (poolside.ai)

Laguna S 2.1 has been unveiled as a powerful new 118 billion parameter Mixture-of-Experts model, designed for long-horizon reasoning tasks and capable of utilizing a colossal context window of up to 1 million tokens. Despite being smaller than its competitors, Laguna S 2.1 excels in performance, scoring 70.2% on the Terminal-Bench 2.1 and maintaining a lead in benchmarks against models significantly larger in scale. This rapid development from training to launch in just nine weeks highlights its efficiency and competitiveness in the AI/ML landscape. This release is particularly significant for the AI community as it demonstrates the potential of compact models to perform complex tasks, including the engineering of a browser engine and solving intricate mathematical problems, such as a proof for Erdős problem #397. Laguna S 2.1's dual thinking modes enhance its problem-solving capabilities, allowing for extended productive reasoning. However, it does come with limitations, such as some challenges in adhering to external tool schemas. Overall, Laguna S 2.1 exemplifies a notable shift in the development of AI models that are both resourceful and capable of achieving results akin to their larger counterparts.