StepFun · 2026-05 · Sparse MoE · Multimodal
Step-3.7-Flash
Step-3.7-Flash is a multi-head attention (MHA) transformer released by StepFun in 2026-05, with 45 layers, hidden size 4096 and a context window of 262,144 tokens.
Layer stack
Attention ×12 Sliding window ×33 · Layers 45
Key facts
| Family | StepFun |
|---|---|
| Released | 2026-05 |
| Params | 198 B |
| Active | 11 B |
| Context | 262,144 tokens |
| Attention | MHA |
| Layers | 45 |
| Hidden | 4,096 |
| Heads | 64 |
| Vocab | 128,896 |
| Position | RoPE |
| Norm | — |
| Activation | — |
| Dtype | bf16 |
| Architecture class | Step3p7ForConditionalGeneration |
Architecture overview
Attention
multi-head attention (MHA). RoPE θ=NaN.
Feed-forward / MoE
It is a sparse mixture-of-experts with 288 routed experts and top-8 routing; about 11B parameters are active per token.
The configuration ships 3 extra multi-token-prediction (MTP) layers used to accelerate decoding.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor Step-3.5-Flash
| Model | 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 |
Most similar architectures
- 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
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
49 fields
| 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 |