StepFun · 2026-03 · MoE sparsa
Step-3.5-Flash
Step-3.5-Flash è un Transformer a attention multi-head (MHA) pubblicato da StepFun nel 2026-03, con 45 strati, dimensione nascosta 4096 e una finestra di contesto di 262,144 token.
Pila di strati
Dati chiave
| Laboratorio | StepFun |
|---|---|
| Pubblicazione | 2026-03 |
| Parametri | 196 mld |
| Attivi | 11 mld |
| Contesto | 262,144 tokens |
| Attention | MHA (64:64) |
| Strati | 45 |
| Dim. nascosta | 4,096 |
| Teste di attention | 64 |
| Vocabolario | 128,896 |
| Codifica posizionale | RoPE |
| Normalizzazione | — |
| Attivazione | — |
| Precisione | bf16 |
| Classe architetturale | Step3p5ForCausalLM |
Panoramica dell'architettura
Attention
attention multi-head (MHA) — 64 q-heads / 64 kv-heads. RoPE θ=NaN.
FFN / MoE
È un mixture-of-experts sparso con 288 esperti instradati e routing top-8; circa 11 milioni di parametri sono attivi per token.
La configurazione include 3 strati extra di predizione multi-token (MTP) usati per accelerare il decodice.
Confronto campo per campo
Confronto con il modello precedente dello stesso laboratorio; in assenza, col modello strutturalmente più vicino. La colonna rapporto = questo modello ÷ modello di confronto.
Confronto completo con il predecessore Step-3
| Modello | Step-3.5-Flash | Step-3 | Rapporto |
|---|---|---|---|
| model_type | step3p5 | step3_vl | ≠ |
| architectures | Step3p5ForCausalLM | Step3VLForConditionalGeneration | ≠ |
| hidden_size | 4096 | 7168 | ×0.571 |
| num_hidden_layers | 45 | 61 | ×0.738 |
| num_attention_heads | 64 | 64 | ≈1 |
| head_dim | 128 | 256 | ×0.500 |
| intermediate_size | 11264 | 18432 | ×0.611 |
| moe_intermediate_size | 1280 | 5120 | ×0.250 |
| rope_theta | 5000000.0×12 + 10000.0×36 | 500000 | ≠ |
| max_position_embeddings | 262144 | — | |
| sliding_window | 512 | — | |
| vocab_size | 128896 | 128815 | ≈1 |
| torch_dtype | bfloat16 | bfloat16 | ＝ |
| num_nextn_predict_layers | 3 | — | |
| yarn_only_types | full_attention×1 | — | |
| max_seq_len | 262144 | 65536 | ×4.00 |
| use_qk_norm | true | — | |
| 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 | 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_groups | 8 | 1 | ×8.00 |
| use_moe | true | — | |
| moe_num_experts | 288 | 48 | ×6.00 |
| moe_top_k | 8 | 3 | ×2.67 |
| share_expert_dim | 1280 | 5120 | ×0.250 |
| moe_layer_offset | 0 | — | |
| moe_every_n_layer | 1 | — | |
| norm_expert_weight | true | false | ≠ |
Architetture più simili
- Step-3.7-Flash (StepFun, 2026-05) 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
I valori di similarità vanno da 0 (nessuna caratteristica categoriale condivisa) a 1 (profili identici).
Campi grezzi del config.json
41 campi
| architectures | Step3p5ForCausalLM |
|---|---|
| model_type | step3p5 |
| yarn_only_types | full_attention×1 |
| hidden_size | 4096 |
| intermediate_size | 11264 |
| num_hidden_layers | 45 |
| max_seq_len | 262144 |
| vocab_size | 128896 |
| torch_dtype | bfloat16 |
| use_qk_norm | true |
| 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 |
| 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 |
| tie_word_embeddings | false |
| 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 |
| num_nextn_predict_layers | 3 |
| 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 | 0.0×47 + 16×1 |
| zero_centered | true |
| max_position_embeddings | 262144 |