System One language models enable faster decision-making

New developments in “System One” language models focus on models designed to output specific decisions rather than generating text word-by-word. These models, such as TypeSafe’s commercial Jev model, function by taking a piece of text and a list of options to return probabilities for each choice in a single pass, resulting in significantly higher speed compared to traditional large language models (LLMs).

Research indicates that any LLM can be converted into a System One model by batching prompts to generate single tokens with structured outputs. In practical applications, such as playing the game Doom, System One versions of models like Qwen3-8B have demonstrated much faster reaction times—making multiple decisions every 190ms compared to roughly one decision every 600ms using standard tool-calling methods. While these models are less flexible than general-purpose LLMs, they offer advantages in speed and consistency for classification and decision-making tasks.