OpenAI · 2020-05 · Dense
GPT-3 175B (paper)
GPT-3 175B (paper) is a multi-head attention (MHA) transformer released by OpenAI in 2020-05, with 96 layers, hidden size 12288 and a context window of 2,048 tokens.
Reference architecture from the original paper; no open weights were released.
Layer stack
Attention ×96 · Layers 96
Key facts
| Family | OpenAI |
|---|---|
| Released | 2020-05 |
| Params | 175 B |
| Context | 2,048 tokens |
| Attention | MHA (96:96) |
| Layers | 96 |
| Hidden | 12,288 |
| Heads | 96 |
| Vocab | 50,257 |
| Position | learned-absolute |
| Norm | — |
| Activation | gelu |
| Dtype | — |
| Architecture class | gpt3-paper |
Architecture overview
Attention
multi-head attention (MHA) — 96 q-heads / 96 kv-heads.
Feed-forward / MoE
It is a dense model: all 175B parameters participate in computing every token.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor GPT-2
| Model | GPT-3 175B (paper) | GPT-2 | Ratio |
|---|---|---|---|
| model_type | gpt3-paper | gpt2 | ≠ |
| architectures | — | GPT2LMHeadModel | |
| hidden_size | 12288 | — | |
| num_attention_heads | 96 | — | |
| max_position_embeddings | 2048 | — | |
| layer_norm_epsilon | — | 0.00001 | |
| vocab_size | 50257 | 50257 | ≈1 |
| _note | GPT-3 未开源权重，本表字段取自 arXiv:2005.14165 Table 2.1 (GPT-3 175B/davinci) | — | |
| source | paper:arXiv:2005.14165 | — | |
| num_layers | 96 | — | |
| n_head_kv | 96 | — | |
| activation | gelu | — | |
| positional_encoding | learned:absolute | — | |
| normalization | LayerNorm(pre+post) | — | |
| residual | post-LN style | — | |
| attention | MHA, alternating local(sp=256)/global per layer (GPT-3 paper §2.1) | — |
Most similar architectures
- GPT-2 (OpenAI, 2019-02) 0.744
- Step-3 (StepFun, 2025-07) 0.687
- Step-3.7-Flash (StepFun, 2026-05) 0.613
- Step-3.5-Flash (StepFun, 2026-03) 0.613
- LLaMA 1 7B (Meta, 2023-02) 0.581
- Qwen1.5 7B (Qwen, 2024-02) 0.528
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
14 fields
| _note | GPT-3 未开源权重，本表字段取自 arXiv:2005.14165 Table 2.1 (GPT-3 175B/davinci) |
|---|---|
| source | paper:arXiv:2005.14165 |
| model_type | gpt3-paper |
| num_layers | 96 |
| hidden_size | 12288 |
| num_attention_heads | 96 |
| n_head_kv | 96 |
| max_position_embeddings | 2048 |
| vocab_size | 50257 |
| activation | gelu |
| positional_encoding | learned:absolute |
| normalization | LayerNorm(pre+post) |
| residual | post-LN style |
| attention | MHA, alternating local(sp=256)/global per layer (GPT-3 paper §2.1) |