Definition

Test-Time Compute

Test-time compute is the third scaling frontier: letting a model think longer at inference time instead of only making it bigger or training it more. Reasoning models such as o1, o3, and DeepSeek R1 generate an internal chain of thought before answering, sometimes for minutes, and outperform larger models on math, code, and logic. For builders it adds a new design axis: how long should this query think?