Qwen3 235B-A22B
Alibaba Cloud (Qwen) · Qwen3 · open weights
- GQA
- RoPE
- Pre-norm
- QK-norm
- MoE
Facts and where they come from
| Released | 2025-04 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 235B | labmodel cardREADME: 235B in total and 22B activated |
| Active parameters | 22B | labmodel cardREADME: 22B activated |
| Context length | 32K tokens | labmodel cardREADME: 32,768 natively |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_moe: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_moe: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | codemodelling codetransformers 5.18.0 qwen3_moe: q_norm and k_norm |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_moe: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
GQA 64q/4kv. 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) | 235B |
|---|---|
| Active per token (modelled) | 22.2B |
| Without embeddings and output head | 234B total, 20.9B active |
| Published weights (Hugging Face count) | 235B |
| KV cache per token, BF16 (layers that grow with context) | 188 KiB |
| KV cache + state at 32K tokens, BF16 | 5.88 GiB |
| Decode FLOPs per token at 4K context | 55.8 GFLOP |
| Prefill FLOPs for a 4K prompt | 197 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 151,936 | 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 | 4 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.full.qk_norm | true | codemodelling codetransformers 5.18.0 qwen3_moe: q_norm and k_norm |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 128 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 1,536 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| layout | 94× full/moe | config.jsonconfig.jsonnum_hidden_layers |
Sources
- config.json @ 8efa617
- model card
- arXiv 2505.09388
- 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 “Qwen3 (235B-A22B)” (name only; see about).