StepFun · 2026-03 · 稀疏 MoE
Step-3.5-Flash
Step-3.5-Flash 是 StepFun 於 2026-03 發布的 多頭注意力（MHA） Transformer，共 45 層，隱藏維度 4096，上下文視窗 262,144 個 token。
層堆疊
注意力 ×12 滑動視窗 ×33 · 層數 45
關鍵參數
| 實驗室 | StepFun |
|---|---|
| 發布時間 | 2026-03 |
| 總參數 | 196 B |
| 激活參數 | 11 B |
| 上下文 | 262,144 tokens |
| 注意力 | MHA (64:64) |
| 層數 | 45 |
| 隱藏維度 | 4,096 |
| 注意力頭 | 64 |
| 詞表 | 128,896 |
| 位置編碼 | RoPE |
| 正規化 | — |
| 激活函數 | — |
| 訓練精度 | bf16 |
| 架構類別 | Step3p5ForCausalLM |
架構概覽
注意力
多頭注意力（MHA） — 64 q-heads / 64 kv-heads. RoPE θ=NaN.
前饋 / MoE
它是稀疏混合專家模型：288 個路由專家、top-8 路由；每個 token 約激活 11B 參數。
設定中包含 3 個額外的多 token 預測（MTP）層，用於加速解碼。
欄位級對比
對比對象為同實驗室的上一代模型；無前代時取結構最相似的模型。倍數欄 = 本模型數值 ÷ 對比模型數值。
與前代 Step-3 的全欄位對比
| 模型 | Step-3.5-Flash | Step-3 | 倍數 |
|---|---|---|---|
| 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 | ≠ |
結構最相似的模型
- 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
相似度取值範圍 0（無共同類別特徵）到 1（結構輪廓完全一致）。
原始 config 欄位
41 個欄位
| 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 |