GPT-3

AI模型历史历史

175 billion parameters and the discovery of in-context learning.

您正在查看一条 AI 模型记录：谁构建了它、它能做什么，以及其背后的硬件与数据。

GPT-3 demonstrated that a large enough language model could perform tasks from a few examples in its prompt — no fine-tuning — launching the API business model and the scaling era. Trained on V100 clusters with 3.1×10²³ FLOP (disclosed), it is the ancestor of every product on this site's Models list.

规格与关键参数

每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。

Overview

Training compute 3.1×10²³ FLOPPublished paper

Training hardware NVIDIA V100 clusters (Microsoft-built)Published paper

它如何与整个技术栈相连

本记录的关联链接一览——点击分组可直接跳转。

开发方

OpenAI

训练于

NVIDIA V100

训练数据包括

Common Crawl

推出时间

Language Models are Few-Shot Learners (2020)

后继产品

GPT-4

前代产品

GPT-2

来源

GPT-3 paper ↗

记录审阅截至 2026-01-15 — 查看方法论.