ArseneLupin v1.1 a JEV-like model (and 2 more) (huggingface.co)

The release of ArseneLupin v1.1 introduces an innovative open decision model that processes various inputs such as documents and JSON records to provide probability distributions for typed questions in a single forward pass. This model features three question types: binary (yes/no), multiple-choice, and ordered scoring. With a robust architecture based on Qwen/Qwen3.5-4B, it comprises 4.66 billion parameters—over 4.2 billion dedicated to decision-making—allowing it to outperform competitors like Jev in several benchmarks. The significance of ArseneLupin v1.1 lies in its accuracy and efficiency in decision-making tasks across different applications, evidenced by its performance on proprietary and public benchmarks. It achieved accuracy rates surpassing those of similar models in scenarios involving untrained decisions, security incident evaluations, and intent classification tasks. Its unique features, including model calibration for specific workflows and compatibility with popular frameworks like transformers, make it a valuable tool for developers and researchers in the AI/ML space looking to enhance their decision-making capabilities.