GPT-4 (Mar 2023)
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- OpenAI
- Organisation type
- Industry
- Country
- United States of America
- Published
- 15 March 2023
- Authors
- OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Bo…
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Multimodal, Language, Vision
- Task
- Language modeling, Language modeling/generation, Question answering, Visual question answering
- Approach
- Self-supervised learning
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 1.8T
- Training data
- 5,416,666,666,667 tokens
- Epochs
- 2
- Batch size
- 60,000,000
Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/
Speculative. Reported secondhand by online sources such as Semianalysis, but not verified by OpenAI. If total number of tokens seen was 13T, text was repeated for 2 epochs, and text was the majority of tokens, then dataset size roughly is 13T*0.75/2 = 4.9T words. Note this examines only the text dataset, since GPT-4 was first and foremost a language model. However, the vision component had its own vision dataset, which we believe accounted for a much smaller part of the compute budget.
not listed
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 2.1 × 10²⁵ FLOP
- How it was established
- Hardware
90% CI: 8.2E+24 to 4.4E+25 NOTE: this is a rough estimate based on public information, much less information than most other systems in the database. Calculation and confidence intervals here: https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9ScslE_nOj5irN9?usp=sharing
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA A100 SXM4 40 GB
- Chips used
- 25,000
- Chip-hours
- 57,000,000
- Wall-clock time
- 2,280 hours (95 days)
- Hardware utilisation
- MFU 34.0%
- Power draw
- 19.9 MW
- Compute cost
- $37,334,305
(Speculative) SemiAnalysis conjectures that GPT-4 training took 90-100 days with utilization of 32-36%.
CANNOT VERIFY, LIKELY HFU, PAYWALLED. (Speculative) SemiAnalysis conjectures that GPT-4 had utilization of 32-36%: https://www.semianalysis.com/p/gpt-4-architecture-infrastructure
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- API access
- Training code
- Unreleased
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Frontier model
- Yes
- Foundation model
- Yes
- Likely above 10²³ FLOP
- Yes
- Why it is tracked
- Highly cited,SOTA improvement,Training cost
- Record confidence
- Likely
- Citations
- 24,490
See the paper, p.1: "On a suite of traditional NLP benchmarks, GPT-4 outperforms both previous large language models and most state-of-the-art systems (which often have benchmark-specific training or hand-engineering)." "On the MMLU benchmark [35, 36], an English-language suite of multiple-choice questions covering 57 subjects, GPT-4 not only outperforms existing models by a considerable margin in English, but also demonstrates strong performance in other languages"
Sources
Where this record came from and when it was last checked.
- Reference
- GPT-4 Technical Report
- Last updated
- 25 May 2026
What the numbers mean
Where it came from
GPT-4 (Mar 2023) was published by OpenAI, in United States of America, in March 2023. The organisation is categorised as industry.
It works in Multimodal, Language, Vision, and is recorded as doing language modeling, Language modeling/generation, Question answering, Visual question answering.
Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.
What went into building it
Training it took roughly 2.1 × 10²⁵ FLOP of computation, on NVIDIA A100 SXM4 40 GB — a measure of what producing the model cost, not of how fast it answers.
Around 5,416,666,666,667 tokens went into training it.
Its inclusion criterion is highly cited,SOTA improvement,Training cost.
Answers
GPT-4 (Mar 2023) — common questions
What is GPT-4 (Mar 2023) used for?
GPT-4 (Mar 2023) works in Multimodal, Language, Vision, and is recorded as handling language modeling, Language modeling/generation, Question answering, Visual question answering. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
How much compute was used to train GPT-4 (Mar 2023)?
Around 2.1 × 10²⁵ FLOP, on NVIDIA A100 SXM4 40 GB. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
What GPU do I need to run GPT-4 (Mar 2023)?
None. GPT-4 (Mar 2023) is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Is GPT-4 (Mar 2023) open source?
No. GPT-4 (Mar 2023) has not had its weights published, so it exists only as a service controlled by its owner.
How many parameters does GPT-4 (Mar 2023) have?
GPT-4 (Mar 2023) has 1.8T parameters. Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
Who created GPT-4 (Mar 2023)?
GPT-4 (Mar 2023) was published by OpenAI, based in United States of America, categorised as industry.
When was GPT-4 (Mar 2023) released?
GPT-4 (Mar 2023) was published in March 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.