Why architecture matters here
Small MoE fails on expert collapse, routing overhead, and unsuitable serving. Architecture matters because SLM economics require careful expert count + routing.
Advertisement
The architecture: every piece explained
The top strip is design. Small dense base. Convert to MoE expert count + top-K. Router gate. Sparse activation.
The middle row is training. Load balance. Distillation from dense teacher. Serving fit small-cluster feasible. Metrics quality per FLOP.
The lower rows are ops. Eval. Failure modes. Ops — sizing + deploy + cost.
Advertisement
End-to-end flow
End-to-end: 1B base + 8 experts + top-2 routing. Distilled from a 7B teacher. Load balance auxiliary loss keeps utilization even. Active params 2B; total 8B. Deployed on modest GPU; quality between 3B and 7B dense.