Abstract

Just last month, a new foundation model called Jev caught the AI world by storm. It is a system one model that doesn't generate text. Rather, it targets decisions and returns categorical probability distributions over predefined options. The early evidence is striking. A benchmark study executed over 346k requests across 37 datasets with Jev scoring 95–99% on standard tasks. It beat open weight models on at least 27 of the 37 datasets. It is weaker on low-resource languages, noisy or fine-grained labels, legal judgments, and rubric scoring. It is cheap, with the evaluation cost at $9.15, and significantly lower latencies than LLMs at 0.36 seconds per request. Another study shows that ordinary open LLMs are already capable Jev-style decision models, with no retraining, and even supports a "Thinking Jev" that reasons on a budget. Meanwhile, more than 60 community Jev-style models have appeared on the public JevBench leaderboard.

About the Speaker

Gaurav Tandon is a VP and Fellow at BILL, where he focuses on building an AI driven financial operating system for SMBs. His past experiences include Orum, Facebook, WeWork, Zulily, Microsoft and Nuance Communications. He received his PhD in Computer Science from Florida Institute of Technology.