StepFun · 2026-05 · Sparsames MoE · Multimodal
Step-3.7-Flash
Step-3.7-Flash ist ein Multi-Head-Attention (MHA)-Transformer, veröffentlicht von StepFun im 2026-05, mit 45 Schichten, Hidden-Größe 4096 und einem Kontextfenster von 262,144 Tokens.
Schichtstapel
Kernfakten
| Labor | StepFun |
|---|---|
| Veröffentlicht | 2026-05 |
| Parameter | 198 Mrd. |
| Aktiv | 11 Mrd. |
| Kontext | 262,144 tokens |
| Attention | MHA |
| Schichten | 45 |
| Hidden-Dim. | 4,096 |
| Attention-Köpfe | 64 |
| Vokabular | 128,896 |
| Positionscodierung | RoPE |
| Normierung | — |
| Aktivierung | — |
| Präzision | bf16 |
| Architekturklasse | Step3p7ForConditionalGeneration |
Architekturüberblick
Attention
Multi-Head-Attention (MHA). RoPE θ=NaN.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 288 gerouteten Experten und Top-8-Routing; etwa 11 Mrd. Parameter sind pro Token aktiv.
Die Konfiguration enthält 3 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 Step-3.5-Flash
| Modell | Step-3.7-Flash | Step-3.5-Flash | Ratio |
|---|---|---|---|
| model_type | step3p7 | step3p5 | ≠ |
| architectures | Step3p7ForConditionalGeneration | Step3p5ForCausalLM | ≠ |
| hidden_size | 4096 | 4096 | ≈1 |
| num_hidden_layers | 45 | 45 | ≈1 |
| num_attention_heads | 64 | 64 | ≈1 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 11264 | 11264 | ≈1 |
| moe_intermediate_size | 1280 | 1280 | ≈1 |
| rope_theta | 5000000.0×12 + 10000.0×36 | 5000000.0×12 + 10000.0×36 | ＝ |
| max_position_embeddings | 262144 | 262144 | ≈1 |
| sliding_window | 512 | 512 | ≈1 |
| vocab_size | 128896 | 128896 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| num_nextn_predict_layers | 3 | 3 | ≈1 |
| im_end_token | <im_end> | — | |
| im_patch_token | <im_patch> | — | |
| im_start_token | <im_start> | — | |
| image_token_len | 169 | — | |
| patch_token_len | 81 | — | |
| image_token_id | 128001 | — | |
| understand_projector_stride | 2 | — | |
| use_im_start_end | true | — | |
| vision_select_layer | -1 | — | |
| projector_bias | false | — | |
| yarn_only_types | full_attention×1 | full_attention×1 | ＝ |
| max_seq_len | 262144 | 262144 | ≈1 |
Strukturell ähnlichste Modelle
- Step-3.5-Flash (StepFun, 2026-03) 1.000
- Step-3 (StepFun, 2025-07) 0.987
- DeepSeekMoE 16B (DeepSeek, 2024-01) 0.796
- MiniMax-M3 (MiniMax, 2026-06) 0.735
- MiniMax-M3 MXFP8 (MiniMax, 2026-06) 0.735
- DeepSeek-V4-Pro (DeepSeek, 2026-04) 0.697
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
49 Felder
| architectures | Step3p7ForConditionalGeneration |
|---|---|
| model_type | step3p7 |
| im_end_token | <im_end> |
| im_patch_token | <im_patch> |
| im_start_token | <im_start> |
| image_token_len | 169 |
| patch_token_len | 81 |
| image_token_id | 128001 |
| understand_projector_stride | 2 |
| use_im_start_end | true |
| vision_select_layer | -1 |
| projector_bias | false |
| yarn_only_types | full_attention×1 |
| hidden_size | 4096 |
| intermediate_size | 11264 |
| num_hidden_layers | 45 |
| max_seq_len | 262144 |
| max_position_embeddings | 262144 |
| vocab_size | 128896 |
| torch_dtype | bfloat16 |
| moe_layers_enum | 3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44 |
| use_mfa | false |
| num_attention_heads | 64 |
| num_attention_groups | 8 |
| head_dim | 128 |
| use_moe | true |
| moe_num_experts | 288 |
| moe_top_k | 8 |
| moe_intermediate_size | 1280 |
| share_expert_dim | 1280 |
| moe_layer_offset | 0 |
| moe_every_n_layer | 1 |
| norm_expert_weight | true |
| moe_router_activation | sigmoid |
| moe_router_scaling_factor | 3 |
| att_impl_type | GQA |
| num_nextn_predict_layers | 3 |
| rope_theta | 5000000.0×12 + 10000.0×36 |
| use_head_wise_attn_gate | true |
| sliding_window | 512 |
| use_moe_router_bias | true |
| need_fp32_gate | true |
| sink | false |
| layer_types | full_attention×12 + sliding_attention×36 |
| partial_rotary_factors | 0.5×12 + 1.0×36 |
| eos_token_id | 1×1 + 2×1 + 128007×1 |
| bos_token_id | 0 |
| swiglu_limits | 0.0×46 + 7×2 |
| swiglu_limits_shared | 0.0×46 + 16×2 |