diff --git a/scripts/cli-entry.js b/scripts/cli-entry.js
index e4916f8dc4..ad0a5ab3a3 100755
--- a/scripts/cli-entry.js
+++ b/scripts/cli-entry.js
@@ -9,29 +9,40 @@
 /**
  * Production bin entry wrapper.
  *
- * Launches dist/cli.js with --expose-gc so that global.gc() is available
- * for the memory-pressure monitor's critical-tier cleanup.
+ * For most commands: launches dist/cli.js with --expose-gc so that
+ * global.gc() is available for the memory-pressure monitor's critical-tier
+ * cleanup.
  *
- * --expose-gc only exposes the function; it has zero runtime cost.
- * global.gc() is called only when RSS hits the critical threshold (0.80),
- * where the 10-200 ms pause is acceptable to avoid an OOM kill.
+ * For `qwen serve`: imports cli.js directly in-process, skipping the
+ * spawnSync overhead (~370ms on EDR-instrumented hosts). The daemon host
+ * process never calls global.gc() — only its ACP children do, and they
+ * independently add --expose-gc via spawnChannel.ts.
  */
 
 import { spawnSync } from 'node:child_process';
-import { fileURLToPath } from 'node:url';
+import { fileURLToPath, pathToFileURL } from 'node:url';
 import { dirname, join } from 'node:path';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const cliPath = join(__dirname, '..', 'dist', 'cli.js');
 
-const result = spawnSync(
-  process.execPath,
-  ['--expose-gc', cliPath, ...process.argv.slice(2)],
-  { stdio: 'inherit' },
-);
-
-if (result.signal) {
-  process.kill(process.pid, result.signal);
-} else {
-  process.exit(result.status ?? 1);
+function isServeCommand() {
+  return process.argv[2] === 'serve';
+}
+
+if (isServeCommand()) {
+  process.argv[1] = cliPath;
+  await import(pathToFileURL(cliPath).href);
+} else {
+  const result = spawnSync(
+    process.execPath,
+    ['--expose-gc', cliPath, ...process.argv.slice(2)],
+    { stdio: 'inherit' },
+  );
+
+  if (result.signal) {
+    process.kill(process.pid, result.signal);
+  } else {
+    process.exit(result.status ?? 1);
+  }
 }
diff --git a/scripts/prepare-package.js b/scripts/prepare-package.js
index 9f25d97d82..588811d2d7 100644
--- a/scripts/prepare-package.js
+++ b/scripts/prepare-package.js
@@ -301,22 +301,31 @@ function writeDistPackageJson(
 
   const cliEntryContent = `#!/usr/bin/env node
 import { spawnSync } from 'node:child_process';
-import { fileURLToPath } from 'node:url';
+import { fileURLToPath, pathToFileURL } from 'node:url';
 import { dirname, join } from 'node:path';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const cliPath = join(__dirname, 'cli.js');
 
-const result = spawnSync(
-  process.execPath,
-  ['--expose-gc', cliPath, ...process.argv.slice(2)],
-  { stdio: 'inherit' },
-);
+function isServeCommand() {
+  return process.argv[2] === 'serve';
+}
 
-if (result.signal) {
-  process.kill(process.pid, result.signal);
+if (isServeCommand()) {
+  process.argv[1] = cliPath;
+  await import(pathToFileURL(cliPath).href);
 } else {
-  process.exit(result.status ?? 1);
+  const result = spawnSync(
+    process.execPath,
+    ['--expose-gc', cliPath, ...process.argv.slice(2)],
+    { stdio: 'inherit' },
+  );
+
+  if (result.signal) {
+    process.kill(process.pid, result.signal);
+  } else {
+    process.exit(result.status ?? 1);
+  }
 }
 `;