diff --git a/core/application/failover.go b/core/application/failover.go
index 3194e294b..f0b5c2c2b 100644
--- a/core/application/failover.go
+++ b/core/application/failover.go
@@ -5,13 +5,28 @@ import (
 	"github.com/mudler/xlog"
 )
 
+// preloadModelByName is a seam over backend.PreloadModelByName so tests can
+// substitute a controllable loader instead of touching real models/disk.
+var preloadModelByName = backend.PreloadModelByName
+
 // applyFailoverWarmTargets pins warm failover targets in the watchdog and
 // loads them, so a switch does not wait for a cold load.
+//
+// SyncPinnedModelsToWatchdog runs synchronously: it is a cheap in-memory
+// update, and the pin must land before the watchdog can evict a target that
+// is about to become (or stay) a chain's active path. Preloading is not
+// cheap — it can download or load a multi-GB model — and this callback runs
+// on the failover manager's single scheduler goroutine (Sync, called from
+// Tick, called from Run), before that tick's probes fire. A slow or hung
+// load here would freeze probing and fail-back for every chain, so it runs
+// in its own goroutine instead of blocking the scheduler loop.
 func (a *Application) applyFailoverWarmTargets(warm []string) {
 	a.SyncPinnedModelsToWatchdog()
-	for _, name := range warm {
-		if _, err := backend.PreloadModelByName(a.ApplicationConfig().Context, a.ModelConfigLoader(), a.ModelLoader(), a.ApplicationConfig(), name); err != nil {
-			xlog.Warn("failover: could not preload warm target", "model", name, "error", err)
+	go func() {
+		for _, name := range warm {
+			if _, err := preloadModelByName(a.ApplicationConfig().Context, a.ModelConfigLoader(), a.ModelLoader(), a.ApplicationConfig(), name); err != nil {
+				xlog.Warn("failover: could not preload warm target", "model", name, "error", err)
+			}
 		}
-	}
+	}()
 }
diff --git a/core/application/failover_test.go b/core/application/failover_test.go
new file mode 100644
index 000000000..a9057425a
--- /dev/null
+++ b/core/application/failover_test.go
@@ -0,0 +1,41 @@
+package application
+
+import (
+	"context"
+	"time"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("applyFailoverWarmTargets", func() {
+	It("returns promptly even when the preload call blocks", func() {
+		// Guards against a regression to a synchronous preload loop: onWarm
+		// runs on the failover manager's single scheduler goroutine, so a
+		// blocking loader here must not block the caller.
+		started := make(chan struct{})
+		release := make(chan struct{})
+		orig := preloadModelByName
+		preloadModelByName = func(ctx context.Context, cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig, name string) ([]string, error) {
+			close(started)
+			<-release // never released within the test's timeout
+			return nil, nil
+		}
+		DeferCleanup(func() { preloadModelByName = orig })
+		DeferCleanup(func() { close(release) })
+
+		app := &Application{applicationConfig: &config.ApplicationConfig{Context: context.Background()}}
+
+		callReturned := make(chan struct{})
+		go func() {
+			defer GinkgoRecover()
+			app.applyFailoverWarmTargets([]string{"warm-a"})
+			close(callReturned)
+		}()
+
+		Eventually(callReturned, time.Second).Should(BeClosed(), "applyFailoverWarmTargets must not wait on the preload goroutine")
+		Eventually(started, time.Second).Should(BeClosed(), "the preload goroutine should still run in the background")
+	})
+})