Nimble takes some text and a schema, and makes typed decisions about the text (github.com)

Nimble, a newly developed AI model, simplifies text processing by making quick, typed decisions based on a given schema, which consists of a series of questions. Each question can either provide a predefined choice or be a true/false query. The model outputs selected answers along with their probabilities, making it particularly efficient as it operates without generating preliminary reasoning—drawing inspiration from TypeSafe's Jev's System One approach. This model is notable for its speed and straightforward implementation, as it was built in just one day, and utilizes a curated dataset of 2,676 examples tailored to specific domains. The significance of Nimble lies in its robust decision-making capabilities, allowing developers to define tasks such as routing requests or checking conditions with greater accuracy. It employs scoring methods that read logits to assess answer probabilities efficiently. Nimble runs on both Apple Silicon and NVIDIA GPU-based systems, necessitating specific hardware configurations for optimal performance. While it currently supports text-only input and has limitations on answer structures and probabilities, the curated training approach enhances its efficacy, positioning Nimble as a promising tool for various AI applications in decision-making contexts.