vLLM Semantic Router introduces Fusion, a new routing primitive that enables multi-model panel serving. Instead of routing to a single model, Fusion fans out a request to a panel of models, uses a judge model to analyze agreement and gaps, and synthesizes one final response. The system is policy-driven: operators configure which requests deserve Fusion (via signals and decisions), which models participate, how failures are handled, and how traces are recorded. Three entry paths are supported — auto routing, explicit Fusion-only routing, and per-request panel overrides. OpenRouter's DRACO benchmark results showing fused panels outperforming individual models are cited as external validation. The design keeps Fusion as a routing algorithm rather than a global default, giving operators fine-grained control over when the added latency is justified.
