TypeSafe Jev Played Chess — And Landed Next to Reasoning Models

Developer Maxim Saplin has integrated TypeSafe’s Jev model into his LLM Chess leaderboard to test its performance against traditional chat-based LLMs. Unlike standard models that generate free-form text, Jev is a 'System One' model designed for structured classification tasks, where it selects from a predefined set of options. By treating chess moves as a classification problem, Saplin successfully enabled Jev to play full games. The results were striking: Jev achieved an Elo rating comparable to mid-tier reasoning models while being significantly faster and cheaper, costing only $0.0015 per game. While Jev struggles with generative tasks like spelling, its efficiency in decision-making and protocol adherence makes it a compelling tool for structured agentic workflows. The experiment highlights that specialized, non-chat models can compete with larger reasoning models in specific, constrained environments, offering a high-performance alternative for developers building automated decision systems.