diff --git a/internal/adapter/converter/unified_converter.go b/internal/adapter/converter/unified_converter.go
index 961ac61..914d493 100644
--- a/internal/adapter/converter/unified_converter.go
+++ b/internal/adapter/converter/unified_converter.go
@@ -74,9 +74,15 @@ func (c *UnifiedConverter) ConvertToFormat(models []*domain.UnifiedModel, filter
 func (c *UnifiedConverter) convertModel(model *domain.UnifiedModel) UnifiedModelData {
 	availability := make([]EndpointStatus, 0, len(model.SourceEndpoints))
 	for _, ep := range model.SourceEndpoints {
+		// Use GetEffectiveState() rather than ep.State directly: the lifecycle
+		// unifier updates the typed ModelState field, while ep.State (the legacy
+		// string) is only set at discovery time and never transitions. Reading
+		// the effective state ensures health-driven transitions surface in the
+		// API response. GetEffectiveState() also normalises legacy string values
+		// ("loaded", "not-loaded", "available") to the typed enum.
 		availability = append(availability, EndpointStatus{
 			Endpoint: ep.EndpointName, // Use endpoint name instead of URL
-			State:    ep.State,
+			State:    string(ep.GetEffectiveState()),
 		})
 	}
 	// OLLA-85: [Unification] Models with different digests fail to unify correctly.
diff --git a/internal/adapter/registry/profile/parsers.go b/internal/adapter/registry/profile/parsers.go
index ea91e53..c2db800 100644
--- a/internal/adapter/registry/profile/parsers.go
+++ b/internal/adapter/registry/profile/parsers.go
@@ -225,6 +225,15 @@ func (p *openAIParser) Parse(data []byte) ([]*domain.ModelInfo, error) {
 			Name:     model.ID,
 			Type:     model.Object, // always "model" in openai responses
 			LastSeen: now,
+			// OpenAI-compatible /v1/models has no size or state fields. For these
+			// backends (vllm, llama.cpp, Infinity, etc.) the presence of a model
+			// in the discovery response IS the availability signal — these servers
+			// only list models that are loaded and ready to serve. Without a
+			// non-zero Size, MapModelState() falls through to "unknown" and the
+			// model never appears available in the unified /olla/models response.
+			// Set a sentinel size of 1 to signal "loaded / available, exact size
+			// not reported by this backend".
+			Size: 1,
 		}
 
 		// openai is stingy with metadata