OpenAI · 2025-08 · Sparse MoE
GPT-OSS 120B
GPT-OSS 120B is a grouped-query attention (GQA) transformer released by OpenAI in 2025-08, with 36 layers, hidden size 2880 and a context window of 131,072 tokens.
Layer stack
Sliding window ×18 Attention ×18 · Layers 36
Key facts
| Family | OpenAI |
|---|---|
| Released | 2025-08 |
| Params | 116.8 B |
| Active | 5.1 B |
| Context | 131,072 tokens |
| Attention | GQA (64:8) |
| Layers | 36 |
| Hidden | 2,880 |
| Heads | 64 |
| Vocab | 201,088 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | MXFP4 |
| Architecture class | GptOssForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 64 q-heads / 8 kv-heads. RoPE θ=150,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 128 routed experts and top-4 routing; about 5.1B parameters are active per 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-3 175B (paper)
| Model | GPT-OSS 120B | GPT-3 175B (paper) | Ratio |
|---|---|---|---|
| model_type | gpt_oss | gpt3-paper | ≠ |
| architectures | GptOssForCausalLM | — | |
| hidden_size | 2880 | 12288 | ×0.234 |
| num_hidden_layers | 36 | — | |
| num_attention_heads | 64 | 96 | ×0.667 |
| num_key_value_heads | 8 | — | |
| head_dim | 64 | — | |
| intermediate_size | 2880 | — | |
| hidden_act | silu | — | |
| num_experts_per_tok | 4 | — | |
| rope_theta | 150000 | — | |
| max_position_embeddings | 131072 | 2048 | ×64.00 |
| sliding_window | 128 | — | |
| rms_norm_eps | 0.00001 | — | |
| vocab_size | 201088 | 50257 | ×4.00 |
| attention_bias | true | — | |
| attention_dropout | 0 | — | |
| eos_token_id | 200002 | — | |
| experts_per_token | 4 | — | |
| initial_context_length | 4096 | — | |
| initializer_range | 0.02 | — | |
| layer_types | sliding_attention×18 + full_attention×18 | — | |
| num_local_experts | 128 | — | |
| output_router_logits | false | — | |
| pad_token_id | 199999 | — | |
| router_aux_loss_coef | 0.9 | — |
Most similar architectures
- GLM-4.5-Air (GLM, 2025-07) 0.999
- Qwen3.6 35B A3B (Qwen, 2026-04) 0.997
- Qwen3 30B A3B (Qwen, 2025-04) 0.997
- Hunyuan-TurboS 30B A3B (Hunyuan, 2026-05) 0.997
- Qwen3.8-Flash-Next (Qwen, 2026-08) 0.997
- Qwen3-Next 80B A3B (Qwen, 2025-09) 0.996
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
30 fields
| architectures | GptOssForCausalLM |
|---|---|
| attention_bias | true |
| attention_dropout | 0 |
| eos_token_id | 200002 |
| experts_per_token | 4 |
| head_dim | 64 |
| hidden_act | silu |
| hidden_size | 2880 |
| initial_context_length | 4096 |
| initializer_range | 0.02 |
| intermediate_size | 2880 |
| layer_types | sliding_attention×18 + full_attention×18 |
| max_position_embeddings | 131072 |
| model_type | gpt_oss |
| num_attention_heads | 64 |
| num_experts_per_tok | 4 |
| num_hidden_layers | 36 |
| num_key_value_heads | 8 |
| num_local_experts | 128 |
| output_router_logits | false |
| pad_token_id | 199999 |
| rms_norm_eps | 0.00001 |
| rope_theta | 150000 |
| router_aux_loss_coef | 0.9 |
| sliding_window | 128 |
| swiglu_limit | 7 |
| tie_word_embeddings | false |
| transformers_version | 4.55.0.dev0 |
| use_cache | true |
| vocab_size | 201088 |