gpt-oss-20b
OpenAI · gpt-oss · open weights
- GQA
- Sliding window
- RoPE
- Pre-norm
- MoE
Facts and where they come from
| Released | 2025-08 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 21B | labmodel cardREADME: 21B parameters with 3.6B active |
| Active parameters | 3.6B | labmodel cardREADME: 3.6B active |
| Context length | 128K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for gpt_oss: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for gpt_oss: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | no | codemodelling codeno q/k normalisation in the attention block |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for gpt_oss: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
12× GQA 64q/8kv, window 128 + 12× GQA 64q/8kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 20.9B |
|---|---|
| Active per token (modelled) | 4.19B |
| Without embeddings and output head | 19.7B total, 3.03B active |
| Published weights (Hugging Face count) | 20.9B (packed low-bit tensors, so not comparable) |
| KV cache per token, BF16 (layers that grow with context) | 24 KiB |
| KV cache + state at 128K tokens, BF16 | 3 GiB |
| Decode FLOPs per token at 4K context | 8.05 GFLOP |
| Prefill FLOPs for a 4K prompt | 26.6 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 2,880 | config.jsonconfig.jsonhidden_size |
| vocab | 201,088 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 64 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 64 | config.jsonconfig.jsonhead_dim |
| mixers.full.bias | true | codemodelling codeattention_bias: true |
| mixers.sliding.type | attn | codemodelling codeattention block |
| mixers.sliding.heads | 64 | config.jsonconfig.jsonnum_attention_heads |
| mixers.sliding.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.sliding.head_dim | 64 | config.jsonconfig.jsonhead_dim |
| mixers.sliding.window | 128 | config.jsonconfig.jsonsliding_window |
| mixers.sliding.bias | true | codemodelling codeattention_bias: true |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 32 | config.jsonconfig.jsonnum_local_experts |
| ffns.moe.active | 4 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 2,880 | config.jsonconfig.jsonintermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| layout | 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe · 1× sliding/moe · 1× full/moe | config.jsonconfig.jsonlayer_types |
Sources
- config.json @ 6cee5e8
- model card
- arXiv 2508.10925
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “GPT-OSS (20B)” (name only; see about).