StepFun · 2025-07 · Sparsames MoE · Multimodal
Step-3
Step-3 ist ein Multi-Head-Attention (MHA)-Transformer, veröffentlicht von StepFun im 2025-07, mit 61 Schichten, Hidden-Größe 7168 und einem Kontextfenster von 65,536 Tokens.
Schichtstapel
Attention ×61 · Schichten 61
Kernfakten
| Labor | StepFun |
|---|---|
| Veröffentlicht | 2025-07 |
| Parameter | 321 Mrd. |
| Aktiv | 38 Mrd. |
| Kontext | 65,536 tokens |
| Attention | MHA |
| Schichten | 61 |
| Hidden-Dim. | 7,168 |
| Attention-Köpfe | 64 |
| Vokabular | 128,815 |
| Positionscodierung | RoPE |
| Normierung | — |
| Aktivierung | — |
| Präzision | bf16 |
| Architekturklasse | Step3VLForConditionalGeneration |
Architekturüberblick
Attention
Multi-Head-Attention (MHA). RoPE θ=500,000.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 48 gerouteten Experten und Top-3-Routing; etwa 38 Mrd. Parameter sind pro Token aktiv.
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 ähnlichsten Modell Step-3.7-Flash
| Modell | Step-3 | Step-3.7-Flash | Ratio |
|---|---|---|---|
| model_type | step3_vl | step3p7 | ≠ |
| architectures | Step3VLForConditionalGeneration | Step3p7ForConditionalGeneration | ≠ |
| hidden_size | 7168 | 4096 | ×1.75 |
| num_hidden_layers | 61 | 45 | ×1.36 |
| num_attention_heads | 64 | 64 | ≈1 |
| head_dim | 256 | 128 | ×2.00 |
| intermediate_size | 18432 | 11264 | ×1.64 |
| moe_intermediate_size | 5120 | 1280 | ×4.00 |
| rope_theta | 500000 | 5000000.0×12 + 10000.0×36 | ≠ |
| max_position_embeddings | — | 262144 | |
| sliding_window | — | 512 | |
| vocab_size | 128815 | 128896 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | = |
| num_nextn_predict_layers | — | 3 | |
| im_end_token | <im_end> | <im_end> | = |
| im_patch_token | <im_patch> | <im_patch> | = |
| im_start_token | <im_start> | <im_start> | = |
| image_token_len | 169 | 169 | ≈1 |
| patch_token_len | 81 | 81 | ≈1 |
| understand_projector_stride | 2 | 2 | ≈1 |
| projector_bias | false | false | = |
| image_token_id | 128001 | 128001 | ≈1 |
| bos_token_id | 0 | 0 | = |
| eos_token_id | 128805 | 1×1 + 2×1 + 128007×1 | ≠ |
| max_seq_len | 65536 | 262144 | ×0.250 |
| max_position_embedding | 65536 | — |
Strukturell ähnlichste Modelle
- Step-3.7-Flash (StepFun, 2026-05) 0.987
- Step-3.5-Flash (StepFun, 2026-03) 0.987
- DeepSeekMoE 16B (DeepSeek, 2024-01) 0.787
- MiniMax-M3 (MiniMax, 2026-06) 0.731
- MiniMax-M3 MXFP8 (MiniMax, 2026-06) 0.731
- Qwen3.8 2.4T A95B (Qwen, 2026-08) 0.697
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
30 Felder
| architectures | Step3VLForConditionalGeneration |
|---|---|
| model_type | step3_vl |
| im_end_token | <im_end> |
| im_patch_token | <im_patch> |
| im_start_token | <im_start> |
| image_token_len | 169 |
| patch_token_len | 81 |
| understand_projector_stride | 2 |
| projector_bias | false |
| image_token_id | 128001 |
| bos_token_id | 0 |
| eos_token_id | 128805 |
| hidden_size | 7168 |
| intermediate_size | 18432 |
| num_hidden_layers | 61 |
| max_seq_len | 65536 |
| max_position_embedding | 65536 |
| vocab_size | 128815 |
| torch_dtype | bfloat16 |
| moe_layers_enum | 4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59 |
| num_attention_heads | 64 |
| num_attention_groups | 1 |
| head_dim | 256 |
| share_q_dim | 2048 |
| moe_num_experts | 48 |
| moe_top_k | 3 |
| moe_intermediate_size | 5120 |
| share_expert_dim | 5120 |
| norm_expert_weight | false |
| rope_theta | 500000 |