diff --git a/Packages/ContinuityKit/Sources/Ingest/PreparationQueue+Stems.swift b/Packages/ContinuityKit/Sources/Ingest/PreparationQueue+Stems.swift
index 8226163..c51a8a9 100644
--- a/Packages/ContinuityKit/Sources/Ingest/PreparationQueue+Stems.swift
+++ b/Packages/ContinuityKit/Sources/Ingest/PreparationQueue+Stems.swift
@@ -73,7 +73,12 @@ extension PreparationQueue {
             // don't spend CPU-minutes redoing them (reconcileStemLinks links them up next pass).
             if StemCache.hasStems(key: key) {
                 await limiter.release()
-                await MainActor.run { _ = self?.stemsInFlight.remove(key) }
+                let queueDrained = await MainActor.run { () -> Bool in
+                    guard let self else { return true }
+                    self.stemsInFlight.remove(key)
+                    return self.stemsInFlight.isEmpty
+                }
+                if queueDrained { OnnxStemSeparator.releaseSession() }
                 return
             }
             do {
@@ -98,7 +103,15 @@ extension PreparationQueue {
             // key we just wrote (it may not be in the protected set if the queue moved on).
             let protected = await MainActor.run { self?.protectedStemKeys ?? [] }
             StemCache.enforceBudget(protecting: protected.union([key]))
-            await MainActor.run { _ = self?.stemsInFlight.remove(key) }
+            let queueDrained = await MainActor.run { () -> Bool in
+                guard let self else { return true }
+                self.stemsInFlight.remove(key)
+                return self.stemsInFlight.isEmpty
+            }
+            // Last separation done: free the ORT session's weights instead of holding them
+            // under live playback for the rest of the process (jetsam margin on device).
+            // The next batch re-pays one model load — an offline job can afford that.
+            if queueDrained { OnnxStemSeparator.releaseSession() }
         }
     }
 }
diff --git a/Packages/ContinuityKit/Sources/Ingest/StemSeparator.swift b/Packages/ContinuityKit/Sources/Ingest/StemSeparator.swift
index 372aa50..f9eea9d 100644
--- a/Packages/ContinuityKit/Sources/Ingest/StemSeparator.swift
+++ b/Packages/ContinuityKit/Sources/Ingest/StemSeparator.swift
@@ -74,6 +74,16 @@ final class OnnxStemSeparator: StemSeparating {
     private static let sessionLock = NSLock()
     nonisolated(unsafe) private static var cachedSession: (path: String, session: ORTSession)?
 
+    /// Drops the cached session, returning its weights/arenas to the OS. Called when the
+    /// separation queue drains: the session is worth caching *between back-to-back tracks*,
+    /// but holding hundreds of MB of transformer weights under live playback + UI for the
+    /// rest of the process lifetime is exactly the jetsam margin long sessions die on.
+    static func releaseSession() {
+        sessionLock.lock()
+        defer { sessionLock.unlock() }
+        cachedSession = nil
+    }
+
     private static func sharedSession(modelURL: URL) throws -> ORTSession {
         sessionLock.lock()
         defer { sessionLock.unlock() }
@@ -86,6 +96,10 @@ final class OnnxStemSeparator: StemSeparating {
             // plenty for an offline cache-once job.
             try options.setIntraOpNumThreads(1)
             try options.addConfigEntry(withKey: "session.intra_op.allow_spinning", value: "0")
+            // Prepacking rewrites weights into kernel-friendly layouts as ADDITIONAL copies —
+            // for an 85M-param transformer that's hundreds of MB of duplicated RSS for a
+            // marginal speedup on an offline cache-once job. Keeping the originals only.
+            try options.addConfigEntry(withKey: "session.disable_prepacking", value: "1")
             let session = try ORTSession(env: env, modelPath: modelURL.path, sessionOptions: options)
             cachedSession = (modelURL.path, session)
             return session