Reasoning model
A model trained to think at length, in a hidden or visible chain of thought, before giving its final answer.
Also: thinking model, large reasoning model
Reasoning models are trained with reinforcement learning to produce long chains of thought, trying approaches, checking work and backtracking, before answering. OpenAI's o1, previewed in September 2024, was the first widely available example, followed by DeepSeek-R1 in January 2025, whose open weights and published method spread the technique quickly, and reasoning modes in Gemini, Claude, Qwen and others. They do far better on competition maths, science and coding, at the cost of longer waits and more tokens per answer.