Gemini 2.5 Flash Failing on Cold Starts

Hello,

I encountered this issue about 10 days ago, using an Angular app with Vertex AI to call gemini-2.5-flash:
const vertexAI = getAI(this.firebaseApp);
this.model = getGenerativeModel(vertexAI, {
model: “gemini-2.5-flash”,
systemInstruction: systemInstruction,
tools: [productsToolSet],
});

I’m wondering if anyone else is experiencing it, and if Google is aware of it and working on a fix. Thanks in advance.

Expected Behavior:
The model should successfully process all tool calls on a cold start, regardless of the function’s parameters or return type, and return a 200 OK status.

-Scenario A (Fails): Tool returns a complex object

  1. Configure a chat session using the gemini-2.5-flash model.
  2. Define a tool with no parameters that returns a complex object, such as getUsers, which returns User array object.
  3. On a cold start, send a prompt to trigger this tool (e.g., “list users”).
  4. Result: The API call fails with a 500 Internal Server Error.

-Scenario B (Succeeds): Tool returns a primitive type

  1. Configure a chat session using the gemini-2.5-flash model.
  2. Define a tool with no parameters that returns a primitive type, such as displayCart, which returns a string.
  3. On a cold start, send a simple text prompt (e.g., “Hello”) to warm up the model.
  4. Send a second prompt to trigger the “displayCart” tool (e.g., “list cart”).
  5. Result: The API call succeeds with a 200 OK status.

Additional Information:
This bug is specific to the gemini-2.5-flash model. The issue is not reproducible with the gemini-2.0-flash-001 model.
Attached screenshot shows the 500 error from the console.

Hello,

Welcome to the Forum!!

500 error is a temparary error and it should resolve itself after time. You mentioned you observed this issue 10 days ago, I hope it is working fine now.

Thanks for your comments. I’ve actually been experiencing this issue for the past 10 days. This morning, I’m still encountering 500 errors during cold starts with ‘gemini-2.5-flash’. I also tried the new model, ‘gemini-2.5-flash-lite’, and it had the same issue on cold starts.

This issue seems to have resolved itself. For the last two days, I no longer need to run two silent prompts in the application to warm up the Gemini 2.5 Flash model. I removed that code, and now the applicacation runs any first prompt, simple or complex, without errors.

Hello,

Just to confirm, are you using firebase?

If yes, I would recommend raising you concern on firebase support, there you will find better help on issue related to firebase.