Testing System One models through their architecture, and inferring Jev's (www.silasdata.com)

On September 15, 2026, TypeSafe AI unveiled Jev alongside a novel class of models known as System One models, designed to enhance decision-making processes by providing probabilities for multiple options in a single forward pass, rather than generating token outputs. This innovation is significant for the AI/ML community as it highlights a shift in model architecture—rather than being tied to a specific classification task, these models leverage a flexible input approach, allowing them to adapt to various decision scenarios without requiring retraining. The rapid emergence of open-source alternatives like Laya, Lev, and CLM underscores a growing interest in this architecture, with each implementation differing in approach yet fundamentally built on pretrained language models. The architectural distinctions among the models affect their performance and behavior significantly. Laya utilizes ModernBERT and processes state and options together, while Lev adapts Qwen models by using different input orders and integrating LoRA adapters, and CLM employs a more isolated processing method with embeddings and cosine similarity for option evaluation. Experiments revealed that while caching embeddings leads to better performance and stability, model accuracy varies widely, particularly with increasing option numbers, as seen with the dramatic drop in CLM’s performance due to its architecture limitations. This emphasis on architectural influence in model evaluation encourages future research to explore implications on scalability, efficiency, and versatility in decision-making tasks within AI systems.