Reasoning Model

Overview

A reasoning model is built to think before it answers. Rather than producing a response in a single pass, it generates a chain of intermediate reasoning, working through a problem step by step much as a person might on paper, before committing to a final answer. This deliberate, extended process markedly improves performance on hard tasks in mathematics, coding, and logic, where a snap answer often fails.

How it works

These models are trained to produce extended reasoning and then spend that capability at inference time, generating many intermediate tokens of internal working for a single question. This is sometimes called inference-time or test-time scaling: instead of improving a model only by making it bigger or training it longer, more compute is spent at the moment of answering, letting the model explore, check, and refine its own reasoning before responding.

Why it matters

Reasoning models change the shape of compute demand. They shift effort from one-time training toward continuous inference, since each answer now involves generating far more tokens than a direct response would. That makes inference throughput and cost per token central concerns, and it means the hardware serving these models must sustain heavy, sustained generation, exactly the regime current-generation accelerators are built for.