Qwen · 2026-08 · Sparse MoE
Qwen3.8 2.4T A95B
Qwen3.8 2.4T A95B is a grouped-query attention (GQA) transformer released by Qwen in 2026-08, with 92 layers, hidden size 8192 and a context window of 262,144 tokens.
Layer stack
Linear ×69 Attention ×23 · Layers 92
Key facts
| Family | Qwen |
|---|---|
| Released | 2026-08 |
| Params | 2,399.6 B |
| Active | 95 B |
| Context | 262,144 tokens |
| Attention | GQA (64:4) |
| Layers | 92 |
| Hidden | 8,192 |
| Heads | 64 |
| Vocab | 248,320 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | Qwen3_5MoeForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 64 q-heads / 4 kv-heads.
Feed-forward / MoE
It is a sparse mixture-of-experts with 512 routed experts and top-10 routing; about 95B parameters are active per token.
The configuration ships 1 extra multi-token-prediction (MTP) layers used to accelerate decoding.
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 Qwen3.6 35B A3B
| Model | Qwen3.8 2.4T A95B | Qwen3.6 35B A3B | Ratio |
|---|---|---|---|
| model_type | qwen3_5_moe_text | qwen3_5_moe | ≠ |
| architectures | Qwen3_5MoeForCausalLM | Qwen3_5MoeForConditionalGeneration | ≠ |
| hidden_size | 8192 | 2048 | ×4.00 |
| num_hidden_layers | 92 | 40 | ×2.30 |
| num_attention_heads | 64 | 16 | ×4.00 |
| num_key_value_heads | 4 | 2 | ×2.00 |
| head_dim | 256 | 256 | ≈1 |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 10 | 8 | ×1.25 |
| moe_intermediate_size | 2048 | 512 | ×4.00 |
| max_position_embeddings | 262144 | 262144 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 248320 | 248320 | ≈1 |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| attn_output_gate | true | true | ＝ |
| bos_token_id | 248044 | 248044 | ≈1 |
| dtype | bfloat16 | bfloat16 | ＝ |
| eos_token_id | 248044 | 248044 | ≈1 |
| full_attention_interval | 4 | 4 | ≈1 |
| initializer_range | 0.02 | 0.02 | ≈1 |
| layer_types | linear_attention×69 + full_attention×23 | linear_attention×30 + full_attention×10 | ≠ |
| linear_conv_kernel_dim | 4 | 4 | ≈1 |
| linear_key_head_dim | 128 | 128 | ≈1 |
| linear_num_key_heads | 16 | 16 | ≈1 |
| linear_num_value_heads | 128 | 32 | ×4.00 |
Most similar architectures
- GLM-4.5 (GLM, 2025-07) 0.994
- GLM-4.6 (GLM, 2025-10) 0.994
- Qwen3 235B A22B (Qwen, 2025-04) 0.991
- MiniMax-Text-01 (MiniMax, 2025-01) 0.982
- MiniMax-M1 (MiniMax, 2025-06) 0.982
- GLM-4.5-Air (GLM, 2025-07) 0.980
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
39 fields
| architectures | Qwen3_5MoeForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| attn_output_gate | true |
| bos_token_id | 248044 |
| dtype | bfloat16 |
| eos_token_id | 248044 |
| full_attention_interval | 4 |
| head_dim | 256 |
| hidden_act | silu |
| hidden_size | 8192 |
| initializer_range | 0.02 |
| layer_types | linear_attention×69 + full_attention×23 |
| linear_conv_kernel_dim | 4 |
| linear_key_head_dim | 128 |
| linear_num_key_heads | 16 |
| linear_num_value_heads | 128 |
| linear_value_head_dim | 128 |
| mamba_ssm_dtype | float32 |
| max_position_embeddings | 262144 |
| model_type | qwen3_5_moe_text |
| moe_intermediate_size | 2048 |
| mtp_num_hidden_layers | 1 |
| mtp_use_dedicated_embeddings | false |
| num_attention_heads | 64 |
| num_experts | 512 |
| num_experts_per_tok | 10 |
| num_hidden_layers | 92 |
| num_key_value_heads | 4 |
| output_gate_type | swish |
| output_router_logits | false |
| partial_rotary_factor | 0.25 |
| rms_norm_eps | 0.000001 |
| router_aux_loss_coef | 0.001 |
| shared_expert_intermediate_size | 2048 |
| tie_word_embeddings | false |
| transformers_version | 4.57.3 |
| use_cache | true |
| vocab_size | 248320 |