GPT-4

OpenAI · March 2023

deprecatedClosedmixture of expertsmultimodalAPI Available
Parameters~1.7T (est. MoE)
Context Window8K/32K tokens
VariantsGPT-4, GPT-4 32K
Sunset DateOctober 2025

Why It Matters

First multimodal frontier model that could process both text and images. Passed the bar exam and scored in the 90th percentile on the SAT, demonstrating expert-level reasoning across many domains.

Description

OpenAI's first multimodal model, able to understand both text and images. Believed to use a mixture-of-experts architecture (where multiple specialized sub-networks collaborate on each response) with an estimated 1.7 trillion parameters. Represented a massive leap in reasoning, coding ability, and factual accuracy over GPT-3.5.

Notable Milestones

  • Passed the Uniform Bar Exam in the 90th percentile
  • Scored 90th percentile on the SAT
  • Powered Bing Chat, Duolingo Max, and Khan Academy's AI tutor Khanmigo
  • First model widely adopted for professional legal, medical, and financial tasks

Benchmark Scores

MMLUMassive Multitask Language Understanding — 57 subjects
86.4%
HumanEvalCode generation pass@1 — Python problems
67.0%

Key Innovations

Multimodal
MultimodalProcessing multiple types of input (text, images, audio, video) in a single model.
MoE
MoEArchitecture where only a fraction of the model's parameters are active for each input, allowing massive scale with lower compute.
Reasoning
ReasoningStructured step-by-step problem solving, often using chain-of-thought or tree-of-thought approaches.

Family Tree

Related Research (5)

2020 · OpenAI

Found that model performance follows power laws in compute, parameters, and data. Provided the mathematical framework for scaling decisions.

Sparse MoEScaling
2017 · Google

Introduced sparsely-gated Mixture-of-Experts layers for scaling model capacity without proportional compute increase.

2021 · Google

Simplified MoE routing to scale to trillions of parameters efficiently. Influenced Mixtral and GPT-4/5 MoE architectures.

ReActReasoning
2022 · Princeton / Google

Combined chain-of-thought reasoning with external tool use (APIs, search), improving QA and decision-making through interleaved reasoning and action.

GPT-4Scaling
2023 · OpenAI

Described GPT-4's multimodal capabilities and performance across professional/academic benchmarks, setting new SOTA on bar exam, MMLU, and many others…

Enabled By

H100NVIDIA · September 2022
1,979 TFLOPS FP8 Tensor

External Links