StepFun · 2025-07 · MoE épars · Multimodal
Step-3
Step-3 est un Transformer à attention multi-tête (MHA) publié par StepFun en 2025-07, avec 61 couches, une dimension cachée de 7168 et une fenêtre de contexte de 65,536 tokens.
Pile de couches
Attention ×61 · Couches 61
Faits clés
| Laboratoire | StepFun |
|---|---|
| Publication | 2025-07 |
| Paramètres | 321 Md |
| Actifs | 38 Md |
| Contexte | 65,536 tokens |
| Attention | MHA |
| Couches | 61 |
| Dimension cachée | 7,168 |
| Têtes d'attention | 64 |
| Vocabulaire | 128,815 |
| Encodage positionnel | RoPE |
| Normalisation | — |
| Activation | — |
| Précision | bf16 |
| Classe d'architecture | Step3VLForConditionalGeneration |
Vue d'ensemble de l'architecture
Attention
attention multi-tête (MHA). RoPE θ=500,000.
FFN / MoE
C’est un mélange d’experts épars : 48 experts routés avec routage top-3 ; environ 38Md de paramètres sont actifs par token.
Comparaison champ à champ
Comparé au modèle précédent du même labo ; à défaut, au modèle structurellement le plus proche. La colonne ratio = ce modèle ÷ le modèle comparé.
Comparaison complète avec le plus proche, Step-3.7-Flash
| Modèle | 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 | — |
Architectures les plus proches
- 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
La similarité varie de 0 (aucune caractéristique catégorielle commune) à 1 (profils identiques).
Champs bruts du config.json
30 champs
| 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 |