Qwen · 2025-04 · Sparse MoE
Qwen3 235B A22B
Qwen3 235B A22B is a grouped-query attention (GQA) transformer released by Qwen in 2025-04, with 94 layers, hidden size 4096 and a context window of 262,144 tokens.
Layer stack
Attention ×94 · Layers 94
Key facts
| Family | Qwen |
|---|---|
| Released | 2025-04 |
| Params | 234.6 B |
| Active | 21.7 B |
| Context | 262,144 tokens |
| Attention | GQA (64:4) |
| Layers | 94 |
| Hidden | 4,096 |
| Heads | 64 |
| Vocab | 151,936 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | Qwen3MoeForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 64 q-heads / 4 kv-heads. RoPE θ=5,000,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 128 routed experts and top-8 routing; about 21.7B 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 Qwen3 0.6B
| Model | Qwen3 235B A22B | Qwen3 0.6B | Ratio |
|---|---|---|---|
| model_type | qwen3_moe | qwen3 | ≠ |
| architectures | Qwen3MoeForCausalLM | Qwen3ForCausalLM | ≠ |
| hidden_size | 4096 | 1024 | ×4.00 |
| num_hidden_layers | 94 | 28 | ×3.36 |
| num_attention_heads | 64 | 16 | ×4.00 |
| num_key_value_heads | 4 | 8 | ×0.500 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 12288 | 3072 | ×4.00 |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 8 | — | |
| moe_intermediate_size | 1536 | — | |
| rope_theta | 5000000 | 1000000 | ×5.00 |
| max_position_embeddings | 262144 | 40960 | ×6.40 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 151936 | 151936 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| bos_token_id | 151643 | 151643 | ≈1 |
| decoder_sparse_step | 1 | — | |
| eos_token_id | 151645 | 151645 | ≈1 |
| initializer_range | 0.02 | 0.02 | ≈1 |
| max_window_layers | 94 | 28 | ×3.36 |
| norm_topk_prob | true | — | |
| num_experts | 128 | — | |
| output_router_logits | false | — |
Most similar architectures
- GLM-4.6 (GLM, 2025-10) 0.999
- GLM-4.5 (GLM, 2025-07) 0.999
- Qwen3.8 2.4T A95B (Qwen, 2026-08) 0.991
- MiniMax-Text-01 (MiniMax, 2025-01) 0.991
- MiniMax-M1 (MiniMax, 2025-06) 0.991
- GLM-4.5-Air (GLM, 2025-07) 0.989
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
31 fields
| architectures | Qwen3MoeForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| bos_token_id | 151643 |
| decoder_sparse_step | 1 |
| eos_token_id | 151645 |
| head_dim | 128 |
| hidden_act | silu |
| hidden_size | 4096 |
| initializer_range | 0.02 |
| intermediate_size | 12288 |
| max_position_embeddings | 262144 |
| max_window_layers | 94 |
| model_type | qwen3_moe |
| moe_intermediate_size | 1536 |
| norm_topk_prob | true |
| num_attention_heads | 64 |
| num_experts | 128 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 94 |
| num_key_value_heads | 4 |
| output_router_logits | false |
| rms_norm_eps | 0.000001 |
| rope_theta | 5000000 |
| router_aux_loss_coef | 0.001 |
| tie_word_embeddings | false |
| torch_dtype | bfloat16 |
| transformers_version | 4.51.0 |
| use_cache | true |
| use_sliding_window | false |
| vocab_size | 151936 |