diff --git a/nemoclaw/src/index.ts b/nemoclaw/src/index.ts
index 4d62108b2c..37def202b6 100644
--- a/nemoclaw/src/index.ts
+++ b/nemoclaw/src/index.ts
@@ -218,7 +218,8 @@ function activeModelEntries(
   onboardCfg: ReturnType<typeof loadOnboardConfig>,
   fallbackModel = "",
 ): ModelProviderEntry[] {
-  const activeModel = onboardCfg?.model ?? fallbackModel;
+  // Prefer fallbackModel (live gateway model) over the potentially stale onboard config (#2608).
+  const activeModel = fallbackModel || onboardCfg?.model || "";
   if (!activeModel) {
     return [
       {
@@ -330,31 +331,28 @@ export default function register(api: OpenClawPluginApi): void {
     handler: (ctx) => handleSlashCommand(ctx, api),
   });
 
-  // 2. Register nvidia-nim provider — use onboard config if available
+  // 2. Register nvidia-nim provider — always probe the live gateway inference
+  // state so the TUI footer reflects the current model after a runtime
+  // `openshell inference set` (#2608).
   const onboardCfg = loadOnboardConfig();
+  const probed = probeOpenShellInference();
 
-  // Prefer onboard config; fall back to live OpenShell inference state when
-  // the config file is unavailable (e.g. inside the sandbox). Only resort to
-  // hardcoded defaults if both lookups fail.
   let bannerEndpoint = onboardCfg ? describeOnboardEndpoint(onboardCfg) : "";
   let bannerProvider = onboardCfg ? describeOnboardProvider(onboardCfg) : "";
-  let bannerModel = onboardCfg?.model ?? "";
-  let probedModel = "";
-
-  if (!bannerEndpoint || !bannerProvider || !bannerModel) {
-    const probed = probeOpenShellInference();
-    if (!bannerEndpoint) bannerEndpoint = probed.endpoint;
-    if (!bannerProvider) bannerProvider = probed.provider;
-    if (!bannerModel) bannerModel = probed.model;
-    probedModel = probed.model;
-  }
+  // Prefer the live gateway model over the stale onboard config model.
+  let bannerModel = probed.model || onboardCfg?.model || "";
+
+  if (!bannerEndpoint) bannerEndpoint = probed.endpoint;
+  if (!bannerProvider) bannerProvider = probed.provider;
 
   if (!bannerEndpoint) bannerEndpoint = "build.nvidia.com";
   if (!bannerProvider) bannerProvider = "NVIDIA Endpoints";
   if (!bannerModel) bannerModel = "nvidia/nemotron-3-super-120b-a12b";
 
   const providerCredentialEnv = onboardCfg?.credentialEnv ?? "NVIDIA_API_KEY";
-  api.registerProvider(registeredProviderForConfig(onboardCfg, providerCredentialEnv, probedModel));
+  api.registerProvider(
+    registeredProviderForConfig(onboardCfg, providerCredentialEnv, probed.model),
+  );
 
   // 3. Register before_tool_call hook to block secrets in memory writes (#1233)
   // NOTE: This relies on OpenClaw's before_tool_call plugin hook contract
diff --git a/nemoclaw/src/register.test.ts b/nemoclaw/src/register.test.ts
index b0c5e864a4..9460e6f7e1 100644
--- a/nemoclaw/src/register.test.ts
+++ b/nemoclaw/src/register.test.ts
@@ -113,6 +113,38 @@ describe("plugin registration", () => {
     expect(logLines.some((line) => line.includes("Model:     llama3.2:latest"))).toBe(true);
   });
 
+  it("prefers live gateway model over stale onboard config model after runtime switch (#2608)", () => {
+    mockedLoadOnboardConfig.mockReturnValue({
+      endpointType: "build",
+      endpointUrl: "https://api.build.nvidia.com/v1",
+      ncpPartner: null,
+      model: "nvidia/nemotron-3-super-120b-a12b",
+      profile: "default",
+      credentialEnv: "NVIDIA_API_KEY",
+      onboardedAt: "2026-03-01T00:00:00.000Z",
+    });
+    mockedExecFileSync.mockReturnValue(
+      JSON.stringify({
+        provider: "NVIDIA",
+        endpoint: "https://api.build.nvidia.com/v1",
+        model: "nvidia/llama-3.3-nemotron-super-49b-v1.5",
+      }),
+    );
+
+    const api = createMockApi();
+    register(api);
+
+    const providerArg = vi.mocked(api.registerProvider).mock.calls[0][0];
+    expect(providerArg.models?.chat).toEqual([
+      expect.objectContaining({ id: "inference/nvidia/llama-3.3-nemotron-super-49b-v1.5" }),
+    ]);
+
+    const logLines = vi.mocked(api.logger.info).mock.calls.map(([message]) => message);
+    expect(
+      logLines.some((line) => line.includes("Model:     nvidia/llama-3.3-nemotron-super-49b-v1.5")),
+    ).toBe(true);
+  });
+
   it("does not treat the provider name as a fallback endpoint", () => {
     mockedExecFileSync.mockReturnValue(
       JSON.stringify({