Production model

GPT-4

Architecture details are not publicly documented.  Per the citation policy, we list only confirmed public facts and refrain from inferring architectural choices.

Architecture

Positional encoding
Norm placement
Norm type
QK-Norm
Activation
Attention
MoE

Architecture details for GPT-4 are not publicly documented. OpenAI’s GPT-4 Technical Report (March 2023) declines to discuss model architecture, training data, training compute, or parameter count. Subsequent statements from OpenAI have not changed this stance.

Per the closed-model policy, this knowledge base does not infer or report architectural claims for undocumented models. The fields below are confirmed public facts only:

Every other architectural slot — positional encoding, normalization placement, activation, attention variant, MoE structure — is recorded as null in this knowledge base’s schema, and the Compare matrix renders these as a dash.

Rumors, leaked details, and third-party reverse-engineering speculation are explicitly out of scope for this knowledge base. See the methodology page for the policy rationale.

Sources

Export

BibTeX
@article{arxiv_2303_08774,
  title         = {GPT-4},
  author        = {OpenAI},
  year          = {2023},
  eprint        = {2303.08774},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2303.08774}
}
CSL JSON
{
  "id": "arxiv_2303_08774",
  "type": "article-journal",
  "title": "GPT-4",
  "author": [
    {
      "literal": "OpenAI"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2023
      ]
    ]
  },
  "URL": "https://arxiv.org/abs/2303.08774",
  "number": "2303.08774",
  "source": "arXiv"
}
RIS
TY  - JOUR
TI  - GPT-4
AU  - OpenAI
PY  - 2023
JO  - arXiv
AN  - arXiv:2303.08774
UR  - https://arxiv.org/abs/2303.08774
ER  -