Step-5-Preview 5-layer slice

A 5-layer slice of stepfun-ai/Step-5-Preview for testing only — not a usable model.

Keeps original decoder layers 0, 3, 4, 88, 91 (one per layer kind: sliding-window + dense MLP; full sparse attention + FP8 MoE; sliding-window + FP8 MoE; sliding-window + BF16 MoE; full sparse attention + dense MLP), renumbered 0-4, plus the vision tower, embeddings and head; MTP layers dropped.

Weights are copied unchanged (FP8 E4M3 128x128 block scales where the original has them).

Downloads last month
-
Safetensors
Model size
24B params
Tensor type
F32
·
BF16
·
F8_E4M3
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pb09204048/Step-5-Preview-5layer

Quantized
(1)
this model