Case study

vLLM Semantic Router

Six milestones in about 14 months from April 2025 to June 2026. This product rapidly evolved from a fixed model choice to full lifecycle management of a Mixture-of-Models system. See the vLLM Blog post, “Beyond a Single Model: Building Mixture-of-Models Systems with vLLM Semantic Router” (2026).

Five-stage evolution diagram: 1. Day 0 Model Choice, 2. IRIS v0.1 Signal to Decision, 3. Athena v0.2 Inference Control, 4. Themis v0.3 Stateful Runtime, 5. MOM Model Lifecycle.
MilestoneWhenWhat changed
IncubationApr 2025Early semantic-routing prototypes began with Mixture-of-Models as the long-term system goal
Initial releaseSep 2025Intent-aware selection between fast and reasoning paths
v0.1 IrisJan 2026Signals, decisions, and route-scoped plugins replaced fixed classification
v0.2 AthenaMar 2026Model selection, memory, RAG, long context, and multimodality expanded routing into an inference control system
v0.3 ThemisJun 2026Stateful routing, projections, replay, protocol support, session continuity, and one production configuration contract made the system operable
Fusion and Micro-AgentJun 2026The router began choosing collaboration patterns, not only individual models

As of mid-2026 the project documents 11 current selection algorithms, from simple nearest-neighbor voting to live metrics-blending — illustrating how much the decision-algorithm menu from earlier in this talk has grown in practice. A few earlier proposals (marked below) didn’t make it into the shipped algorithm set, but are kept here since they still map cleanly onto that taxonomy.

FamilyMethodWhere it fits in the taxonomyWhat it does
ML-basedKNNClustering · offlineFinds similar historical queries and lets nearby examples vote for the best model.
ML-basedKMeansClustering · new models cheapClusters requests and assigns models based on cluster-level quality and efficiency patterns.
ML-basedSVMDifficulty classifier · offlineLearns nonlinear decision boundaries between model preferences using an RBF classifier.
ML-basedMLPDifficulty classifier · offlineUses a neural router to predict the best model from embeddings, with efficient inference through Candle.
AdvancedStaticNo routing — fixed choiceUses a fixed default model or declared order when predictability matters more than adaptation.
AdvancedLatency-AwareMetrics — latency in objectiveSelects the fastest candidate from TPOT and TTFT percentile data when latency budgets dominate.
AdvancedMulti FactorMetrics · multi-signalBalances quality, latency, cost, and load at once, with optional SLO filters.
AdvancedElo *User preference · onlineLearns from user feedback and pairwise preferences using Bradley-Terry style rating updates.
AdvancedRouterDCClustering / similarity · offlineMatches queries to model descriptions with dual-contrastive embedding similarity.
AdvancedAutoMixUncertainty · cascade escalationStarts with cheaper models and escalates based on self-verification to balance cost and quality.
AdvancedPromptDifficulty / capability · LLM-judgedLets a bounded helper model choose directly from the declared candidate models.
AdvancedHybridCombined blocks · metrics-weightedBlends multiple selector scores such as quality, similarity, and cost with configurable weights.
AdvancedThompson Sampling *RL (bandit) · onlineBalances exploration and exploitation online so routing can keep learning while serving production traffic.
AdvancedGMTRouter *User preference · multi-turn statePersonalizes model choice from multi-turn interaction history with graph-based routing.
AdvancedRouter-R1 *RL · multi-round stateUses an external router model to reason about the request before choosing a downstream model.

* No longer documented as a current vLLM Semantic Router algorithm — kept here to illustrate the taxonomy.

Beyond single-model selection, the project also ships five “Looper” algorithms that orchestrate multiple models per request instead of picking just one:

MethodWhere it fits in the taxonomyWhat it does
ConfidenceWorkflow = cascade with escalation policyEscalates sequentially through models until confidence clears a threshold.
RatingsWorkflow = parallel + aggregateGets one response from each candidate with bounded concurrency, similar to consensus voting.
ReMoMRL / multi-round · reasoning synthesisExplores several reasoning paths across multiple rounds, then synthesizes a final answer.
FusionWorkflow = parallel + aggregate, panel + judgeRuns an analysis panel across models, then a judge/synthesis pass merges the results.
WorkflowsWorkflow = planner + multi-stepExecutes a bounded static or planner-generated multi-worker flow.