Qwen · 2026-08 · Sparsames MoE · Multimodal
Qwen3.8-Flash-Next
Qwen3.8-Flash-Next ist ein Grouped-Query-Attention (GQA)-Transformer, veröffentlicht von Qwen im 2026-08, mit 48 Schichten, Hidden-Größe 2560 und einem Kontextfenster von 262,144 Tokens.
Schichtstapel
Kernfakten
| Labor | Qwen |
|---|---|
| Veröffentlicht | 2026-08 |
| Parameter | 123.6 Mrd. |
| Aktiv | 5.1 Mrd. |
| Kontext | 262,144 tokens |
| Attention | GQA (24:2) |
| Schichten | 48 |
| Hidden-Dim. | 2,560 |
| Attention-Köpfe | 24 |
| Vokabular | 248,320 |
| Positionscodierung | RoPE |
| Normierung | RMSNorm |
| Aktivierung | silu |
| Präzision | bf16 |
| Architekturklasse | Qwen4ExpForConditionalGeneration |
Architekturüberblick
Attention
Grouped-Query-Attention (GQA) — 24 q-heads / 2 kv-heads.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 512 gerouteten Experten und Top-10-Routing; etwa 5.1 Mrd. Parameter sind pro Token aktiv.
Die Konfiguration enthält 1 zusätzliche Multi-Token-Prediction-Schichten (MTP) zur Beschleunigung der Dekodierung.
Feldweiser Vergleich
Verglichen mit dem vorherigen Modell desselben Labors; fehlt dieses, mit dem strukturell nächsten Modell. Die Spalte Ratio = dieses Modell ÷ Vergleichsmodell.
Vollfeldvergleich mit dem Vorgänger Qwen3.8 27B
| Modell | Qwen3.8-Flash-Next | Qwen3.8 27B | Ratio |
|---|---|---|---|
| model_type | qwen4_exp | qwen3_5 | ≠ |
| architectures | Qwen4ExpForConditionalGeneration | Qwen3_5ForConditionalGeneration | ≠ |
| hidden_size | 2560 | 5120 | ×0.500 |
| num_hidden_layers | 48 | 64 | ×0.750 |
| num_attention_heads | 24 | 24 | ≈1 |
| num_key_value_heads | 2 | 4 | ×0.500 |
| head_dim | 256 | 256 | ≈1 |
| intermediate_size | — | 17408 | |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 10 | — | |
| moe_intermediate_size | 640 | — | |
| max_position_embeddings | 262144 | 262144 | ≈1 |
| rms_norm_eps | 0.000001 | 0.000001 | ≈1 |
| vocab_size | 248320 | 248320 | ≈1 |
| image_token_id | 248056 | 248056 | ≈1 |
| language_model_only | false | false | ＝ |
| tie_word_embeddings | false | false | ＝ |
| transformers_version | 5.8.0.dev0 | 5.8.0.dev0 | ＝ |
| video_token_id | 248057 | 248057 | ≈1 |
| vision_end_token_id | 248054 | 248054 | ≈1 |
| vision_start_token_id | 248053 | 248053 | ≈1 |
| attention_bias | false | false | ＝ |
| attention_dropout | 0 | 0 | ＝ |
| bos_token_id | 248044 | 248044 | ≈1 |
| dtype | bfloat16 | bfloat16 | ＝ |
| eos_token_id | 248044 | 248044 | ≈1 |
Strukturell ähnlichste Modelle
- Qwen3-Next 80B A3B (Qwen, 2025-09) 1.000
- Qwen3-Coder-Next (Qwen, 2026-01) 1.000
- Qwen3.6 35B A3B (Qwen, 2026-04) 0.998
- GPT-OSS 120B (OpenAI, 2025-08) 0.997
- GLM-4.5-Air (GLM, 2025-07) 0.997
- Qwen3 30B A3B (Qwen, 2025-04) 0.996
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
58 Felder
| architectures | Qwen4ExpForConditionalGeneration |
|---|---|
| image_token_id | 248056 |
| language_model_only | false |
| model_type | qwen4_exp |
| tie_word_embeddings | false |
| transformers_version | 5.8.0.dev0 |
| video_token_id | 248057 |
| vision_end_token_id | 248054 |
| vision_start_token_id | 248053 |
| attention_bias | false |
| attention_dropout | 0 |
| bos_token_id | 248044 |
| dtype | bfloat16 |
| eos_token_id | 248044 |
| full_attention_interval | 4 |
| hc_count | 4 |
| hc_lowrank | 320 |
| head_dim | 256 |
| heads_per_ngram | 8 |
| hidden_act | silu |
| hidden_size | 2560 |
| indexer_budget | 2048 |
| indexer_compress_ratio | 4 |
| indexer_head_dim | 128 |
| indexer_kv_heads | 1 |
| indexer_n_heads | 4 |
| initializer_range | 0.02 |
| layer_types | linear_attention×36 + full_attention×12 |
| linear_conv_kernel_dim | 4 |
| linear_key_head_dim | 128 |
| linear_num_key_heads | 16 |
| linear_num_value_heads | 48 |
| linear_value_head_dim | 128 |
| make_ngram_vocab_size_divisible_by | 128 |
| mamba_ssm_dtype | float32 |
| max_position_embeddings | 262144 |
| moe_intermediate_size | 640 |
| mtp_num_hidden_layers | 1 |
| mtp_use_dedicated_embeddings | false |
| ngram_size | 3 |
| ngram_vocab_size_base | 20000000 |
| num_attention_heads | 24 |
| num_experts | 512 |
| num_experts_per_tok | 10 |
| num_hidden_layers | 48 |
| num_key_value_heads | 2 |
| output_gate_type | sigmoid |
| output_router_logits | false |
| partial_rotary_factor | 0.25 |
| ple_conv_kernel_size | 4 |
| ple_embed_dim | 2560 |
| ple_layer_ids | 2×1 |
| rms_norm_eps | 0.000001 |
| router_aux_loss_coef | 0.001 |
| shared_expert_intermediate_size | 640 |
| split_ngram_parts | 128 |
| use_cache | true |
| vocab_size | 248320 |