[GH-ISSUE #22971] issue: Gemini models stuck in response when using "native" function calling mode with tool-related questions #90597

New Issue

GiteaMirror · 2026-05-15T15:51:56-05:00

GiteaMirror commented

2026-05-15 15:51:56 -05:00

Originally created by @silenceroom on GitHub (Mar 24, 2026).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/22971

Check Existing Issues

I have searched for any existing and/or related issues.
I have searched for any existing and/or related discussions.
I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!).
I am using the latest version of Open WebUI.

Installation Method

Docker

Open WebUI Version

v0.8.10

Ollama Version (if applicable)

No response

Operating System

MacOS

Browser (if applicable)

Chrome

Confirmation

I have read and followed all instructions in README.md.
I am using the latest version of both Open WebUI and Ollama.
I have included the browser console logs.
I have included the Docker container logs.
I have provided every relevant configuration, setting, and environment variable used in my setup.
I have clearly listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
I have documented step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation. My steps:
Start with the initial platform/version/OS and dependencies used,
Specify exact install/launch/configure commands,
List URLs visited, user input (incl. example values/emails/passwords if needed),
Describe all options and toggles enabled or changed,
Include any files or environmental changes,
Identify the expected and actual result at each stage,
Ensure any reasonably skilled user can follow and hit the same issue.

Expected Behavior

Gemini models should respond normally to questions that trigger function calling (tools), even in follow-up messages within a session.

Actual Behavior

When using "native" function calling mode:

First message - works correctly
Second message that triggers function calling (e.g., asking "what time is it now?" which requires a tool) - model gets stuck in responding mode and never completes
Second message that does NOT involve function calling - works normally

Steps to Reproduce

Configure a Gemini model (e.g., Gemini-3-Pro or Gemini-3-Flash) with function calling mode set to "native" via official connection.
Start a new chat session
Send a first message - works correctly
Ask a question that requires tool/function calling (e.g., "what time is it now?")
Model response gets stuck in responding mode and never completes

Logs & Screenshots

Gemini stuck in response mode:

Additional Information

Function calling mode: "native"
Models affected: Gemini-3-Pro, Gemini-3-Flash (possibly other Gemini models)
Behavior differs with non-function-call questions: No issue when 2nd question doesn't require tool calling
When model gets stuck, all following request will be stuck too even without function call request.

Originally created by @silenceroom on GitHub (Mar 24, 2026). Original GitHub issue: https://github.com/open-webui/open-webui/issues/22971 ### Check Existing Issues - [x] I have searched for any existing and/or related issues. - [x] I have searched for any existing and/or related discussions. - [x] I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!). - [x] I am using the latest version of Open WebUI. ### Installation Method Docker ### Open WebUI Version v0.8.10 ### Ollama Version (if applicable) _No response_ ### Operating System MacOS ### Browser (if applicable) Chrome ### Confirmation - [x] I have read and followed all instructions in `README.md`. - [x] I am using the latest version of **both** Open WebUI and Ollama. - [x] I have included the browser console logs. - [x] I have included the Docker container logs. - [x] I have **provided every relevant configuration, setting, and environment variable used in my setup.** - [x] I have clearly **listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup** (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc). - [x] I have documented **step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation**. My steps: - Start with the initial platform/version/OS and dependencies used, - Specify exact install/launch/configure commands, - List URLs visited, user input (incl. example values/emails/passwords if needed), - Describe all options and toggles enabled or changed, - Include any files or environmental changes, - Identify the expected and actual result at each stage, - Ensure any reasonably skilled user can follow and hit the same issue. ### Expected Behavior Gemini models should respond normally to questions that trigger function calling (tools), even in follow-up messages within a session. ### Actual Behavior When using "native" function calling mode: - First message - works correctly - Second message that triggers function calling (e.g., asking "what time is it now?" which requires a tool) - model gets stuck in responding mode and never completes - Second message that does NOT involve function calling - works normally ### Steps to Reproduce 1. Configure a Gemini model (e.g., Gemini-3-Pro or Gemini-3-Flash) with function calling mode set to "native" via official connection. 2. Start a new chat session 3. Send a first message - works correctly 4. Ask a question that requires tool/function calling (e.g., "what time is it now?") 5. Model response gets stuck in responding mode and never completes ### Logs & Screenshots Gemini stuck in response mode: <img width="1108" height="322" alt="Image" src="https://github.com/user-attachments/assets/96bea980-0b37-4219-be80-6a7f4e7d6829" /> ### Additional Information - Function calling mode: "native" - Models affected: Gemini-3-Pro, Gemini-3-Flash (possibly other Gemini models) - Behavior differs with non-function-call questions: No issue when 2nd question doesn't require tool calling - When model gets stuck, all following request will be stuck too even without function call request.

GiteaMirror added the bug label 2026-05-15 15:51:56 -05:00

GiteaMirror commented

2026-05-15 15:51:58 -05:00

@tjbck commented on GitHub (Mar 24, 2026):

Backend logs? We're unable to assist without any concrete reproduction steps.

@tjbck commented on GitHub (Mar 24, 2026): Backend logs? We're unable to assist without any concrete reproduction steps.

GiteaMirror commented

2026-05-15 15:51:59 -05:00

@seppel123 commented on GitHub (Mar 25, 2026):

Looks like i have a similar problem but i use OSS-120B.

@seppel123 commented on GitHub (Mar 25, 2026): Looks like i have a similar problem but i use OSS-120B.

GiteaMirror commented

2026-05-15 15:51:59 -05:00

@soreikomori commented on GitHub (Mar 26, 2026):

I've had this problem for a few months, as described in #19760. Docker logs show essentially nothing:

2026-03-26 22:48:35.613 | INFO     | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/v1/chats/43f2498b-a310-45a4-bd35-cae84d146db9 HTTP/1.1" 200
2026-03-26 22:48:35.648 | INFO     | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/chat/completions HTTP/1.1" 200
2026-03-26 22:48:36.913 | INFO     | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/chat/completed HTTP/1.1" 200
2026-03-26 22:48:36.960 | INFO     | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/v1/chats/43f2498b-a310-45a4-bd35-cae84d146db9 HTTP/1.1" 200
2026-03-26 22:48:36.986 | INFO     | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200

In my particular case, depending on what I set the model up with, I either get an instant "completed" with no content at all- this is the second message in the "completed" response JSON:

{
	"1": {
		"id": "a34efd3b-175e-454e-923e-8c0d9c033a96",
		"role": "assistant",
		"content": "",
		"timestamp": 1774564888
	}
}

Or the eternal loop that @silenceroom described above.

This is the JSON for the "completions" response:

{"status":true,"task_id":"2e9892e6-004d-468e-8279-5b6bf392d586"}

The reproduction steps are pretty much the ones mentioned:

Set up a Google model with the https://generativelanguage.googleapis.com/v1beta/openai URL, Google AI Studio API Key, and OpenAI as provider type. API completions.
Enable native tool calling for any of the resulting models (in my case, gemini-3-flash-preview).
Enable default features for said model and make sure they're functional and enabled (web search, for example).
Make a new chat with said model, and ask a question such as "Give me the latest news for the USA."

That's when it just completes and does nothing else or gets "stuck" thinking- refreshing the page results in the empty response.

@soreikomori commented on GitHub (Mar 26, 2026): I've had this problem for a few months, as described in #19760. Docker logs show essentially nothing: ``` 2026-03-26 22:48:35.613 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/v1/chats/43f2498b-a310-45a4-bd35-cae84d146db9 HTTP/1.1" 200 2026-03-26 22:48:35.648 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/chat/completions HTTP/1.1" 200 2026-03-26 22:48:36.913 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/chat/completed HTTP/1.1" 200 2026-03-26 22:48:36.960 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "POST /api/v1/chats/43f2498b-a310-45a4-bd35-cae84d146db9 HTTP/1.1" 200 2026-03-26 22:48:36.986 | INFO | uvicorn.protocols.http.httptools_impl:send:483 - IP:0 - "GET /api/v1/chats/?page=1 HTTP/1.1" 200 ``` In my particular case, depending on what I set the model up with, I either get an instant "completed" with no content at all- this is the second message in the "completed" response JSON: ``` { "1": { "id": "a34efd3b-175e-454e-923e-8c0d9c033a96", "role": "assistant", "content": "", "timestamp": 1774564888 } } ``` Or the eternal loop that @silenceroom described above. This is the JSON for the "completions" response: ``` {"status":true,"task_id":"2e9892e6-004d-468e-8279-5b6bf392d586"} ``` The reproduction steps are pretty much the ones mentioned: 1. Set up a Google model with the `https://generativelanguage.googleapis.com/v1beta/openai` URL, Google AI Studio API Key, and OpenAI as provider type. API completions. 2. Enable native tool calling for any of the resulting models (in my case, gemini-3-flash-preview). 3. Enable default features for said model and make sure they're functional and enabled (web search, for example). 4. Make a new chat with said model, and ask a question such as "Give me the latest news for the USA." That's when it just completes and does nothing else or gets "stuck" thinking- refreshing the page results in the empty response.

GiteaMirror commented

2026-05-15 15:52:00 -05:00

@Global-Citizen commented on GitHub (Mar 27, 2026):

Same problem for a long time. Function calling doesn't work for Gemini models, but it works for others. I raised this issue under #19397, but unlucky.

@Global-Citizen commented on GitHub (Mar 27, 2026): Same problem for a long time. Function calling doesn't work for Gemini models, but it works for others. I raised this issue under [#19397](https://github.com/open-webui/open-webui/issues/19397), but unlucky.

GiteaMirror commented

2026-05-15 15:52:01 -05:00

@IamGrug commented on GitHub (Mar 27, 2026):

same issue here.

@IamGrug commented on GitHub (Mar 27, 2026): same issue here. <img width="1229" height="656" alt="Image" src="https://github.com/user-attachments/assets/6c2f7c5a-aa98-4013-9a98-a8eb5fef7a5e" />

GiteaMirror commented

2026-05-15 15:52:02 -05:00

@seppel123 commented on GitHub (Mar 27, 2026):

Its not only Gemini Models, i have it with OSS120B, Anthropic Sonnet 4.6 and GPT 5.1 also.
There are no Errors in the log.

@seppel123 commented on GitHub (Mar 27, 2026): Its not only Gemini Models, i have it with OSS120B, Anthropic Sonnet 4.6 and GPT 5.1 also. There are no Errors in the log.

GiteaMirror commented

2026-05-15 15:52:03 -05:00

@Classic298 commented on GitHub (Mar 27, 2026):

I too cannot reproduce with any model in multiple different setups. How can i reproduce this?

@Classic298 commented on GitHub (Mar 27, 2026): I too cannot reproduce with any model in multiple different setups. How can i reproduce this?

GiteaMirror commented

2026-05-15 15:52:03 -05:00

@soreikomori commented on GitHub (Mar 27, 2026):

I too cannot reproduce with any model in multiple different setups. How can i reproduce this?

Did the model answer a prompt just fine? Part of me thinks this could be an API Key issue of some sort... I don't have this problem with GPT, but Gemini (I'm on Paid Tier 1) does indeed just decide to not respond.

@soreikomori commented on GitHub (Mar 27, 2026): > I too cannot reproduce with any model in multiple different setups. How can i reproduce this? Did the model answer a prompt just fine? Part of me thinks this could be an API Key issue of some sort... I don't have this problem with GPT, but Gemini (I'm on Paid Tier 1) does indeed just decide to not respond.

GiteaMirror commented

2026-05-15 15:52:04 -05:00

@Classic298 commented on GitHub (Mar 27, 2026):

I use all models. Gemini, GPT, Claude, Mistral, various providers, never see anything like this. Always works

@Classic298 commented on GitHub (Mar 27, 2026): I use all models. Gemini, GPT, Claude, Mistral, various providers, never see anything like this. Always works

GiteaMirror commented

2026-05-15 15:52:05 -05:00

@soreikomori commented on GitHub (Mar 28, 2026):

What tools have you enabled, @Classic298? Have you set them up?

@soreikomori commented on GitHub (Mar 28, 2026): What tools have you enabled, @Classic298? Have you set them up?

GiteaMirror commented

2026-05-15 15:52:06 -05:00

@soreikomori commented on GitHub (Mar 28, 2026):

Here's the video of what I did and how it keeps, well, not working:

https://github.com/user-attachments/assets/c718f263-7ec7-4968-9b14-b58c5fa086c7

@soreikomori commented on GitHub (Mar 28, 2026): Here's the video of what I did and how it keeps, well, not working: https://github.com/user-attachments/assets/c718f263-7ec7-4968-9b14-b58c5fa086c7

GiteaMirror commented

2026-05-15 15:52:07 -05:00

@Global-Citizen commented on GitHub (Mar 28, 2026):

Hi @Classic298 , these settings work for GLM or Kimi on OpenRouter, but I’ve never been able to get them to work with Gemini on Google provider.

@Global-Citizen commented on GitHub (Mar 28, 2026): ![Image](https://github.com/user-attachments/assets/e618400f-369e-4938-b695-2639869f2c27) ![Image](https://github.com/user-attachments/assets/13905469-6886-46cb-beab-6ae1d4b3ab64) ![Image](https://github.com/user-attachments/assets/77234055-5c58-4598-a68d-33e8177a4c08) Hi @Classic298 , these settings work for GLM or Kimi on OpenRouter, but I’ve never been able to get them to work with Gemini on Google provider.

GiteaMirror commented

2026-05-15 15:52:09 -05:00

@seppel123 commented on GitHub (Apr 1, 2026):

Now i figured something out: my problem is that the reply is sometimes sent to the ‘Thoughts’ section. So if nothing appears as output, I can open the ‘Thoughts’ tag and see the reply. I think something is causing issues with closing and opening the tags for the different sections.
This is definitely a bug.

@seppel123 commented on GitHub (Apr 1, 2026): Now i figured something out: my problem is that the reply is sometimes sent to the ‘Thoughts’ section. So if nothing appears as output, I can open the ‘Thoughts’ tag and see the reply. I think something is causing issues with closing and opening the tags for the different sections. This is definitely a bug.

GiteaMirror commented

2026-05-15 15:52:10 -05:00

@Global-Citizen commented on GitHub (Apr 2, 2026):

Now i figured something out: my problem is that the reply is sometimes sent to the ‘Thoughts’ section. So if nothing appears as output, I can open the ‘Thoughts’ tag and see the reply. I think something is causing issues with closing and opening the tags for the different sections. This is definitely a bug.

Hi @seppel123 , how did you open the ‘Thoughts’ tag?

@Global-Citizen commented on GitHub (Apr 2, 2026): > Now i figured something out: my problem is that the reply is sometimes sent to the ‘Thoughts’ section. So if nothing appears as output, I can open the ‘Thoughts’ tag and see the reply. I think something is causing issues with closing and opening the tags for the different sections. This is definitely a bug. Hi @seppel123 , how did you open the ‘Thoughts’ tag?

GiteaMirror commented

2026-05-15 15:52:10 -05:00

@seppel123 commented on GitHub (Apr 2, 2026):

Just click on the arrow

I hope they can fix that, its really annoying.

@seppel123 commented on GitHub (Apr 2, 2026): Just click on the arrow <img width="383" height="105" alt="Image" src="https://github.com/user-attachments/assets/47390e1b-4d3d-4a3a-8f51-3f2e1a53db66" /> I hope they can fix that, its really annoying.

GiteaMirror commented

2026-05-15 15:52:11 -05:00

@Global-Citizen commented on GitHub (Apr 2, 2026):

Just click on the arrow

I hope they can fix that, its really annoying.

Ok, thanks. Looks like my model didn't even start thinking. No 'Thought' section :/

@Global-Citizen commented on GitHub (Apr 2, 2026): > Just click on the arrow > <img alt="Image" width="383" height="105" src="https://private-user-images.githubusercontent.com/229904247/572905576-47390e1b-4d3d-4a3a-8f51-3f2e1a53db66.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3NzUxMTIzODUsIm5iZiI6MTc3NTExMjA4NSwicGF0aCI6Ii8yMjk5MDQyNDcvNTcyOTA1NTc2LTQ3MzkwZTFiLTRkM2QtNGEzYS04ZjUxLTNmMmUxYTUzZGI2Ni5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjYwNDAyJTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI2MDQwMlQwNjQxMjVaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0wMzE3M2U4YzI5YWI0YjRhYzUxY2I2ODEwMDdlMWM2ZDY2MjMxM2ZkMjJjOGMwNzI3NGU5YTgzNzgyMDI0ZTRhJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCJ9.osNESJtp01rnu8RhJ6hCwEXT4QkG9TqvFNikh7i6d6M"> > > I hope they can fix that, its really annoying. Ok, thanks. Looks like my model didn't even start thinking. No 'Thought' section :/

GiteaMirror commented

2026-05-15 15:52:12 -05:00

@seppel123 commented on GitHub (Apr 2, 2026):

Ok i will create a new Issue for that.

@seppel123 commented on GitHub (Apr 2, 2026): Ok i will create a new Issue for that.

GiteaMirror commented

2026-05-15 15:52:13 -05:00

@ellisd4 commented on GitHub (Apr 3, 2026):

Its not only Gemini Models, i have it with OSS120B, Anthropic Sonnet 4.6 and GPT 5.1 also.
There are no Errors in the log.

I have the same issue with gemma4:31b (via Google API) and gpt-oss:20b (local via ollama).

@ellisd4 commented on GitHub (Apr 3, 2026): > Its not only Gemini Models, i have it with OSS120B, Anthropic Sonnet 4.6 and GPT 5.1 also. > There are no Errors in the log. I have the same issue with gemma4:31b (via Google API) and gpt-oss:20b (local via ollama).

GiteaMirror commented

2026-05-15 15:52:14 -05:00

@trinhkvo commented on GitHub (Apr 5, 2026):

This has been my issue with gemini api for months. The video posted by soreikomori is exactly my observation when prompting any gemini model to use any tool calling.

@trinhkvo commented on GitHub (Apr 5, 2026): This has been my issue with gemini api for months. The video posted by [soreikomori](https://github.com/soreikomori) is exactly my observation when prompting any gemini model to use any tool calling.

GiteaMirror commented

2026-05-15 15:52:14 -05:00

@silenceroom commented on GitHub (Apr 10, 2026):

sorry for the late response, super busy in the last few weeks.

With multiple rounds of test, I can reproduce the issue with below conditions:

Gemini works nomral without "built-in" tools enabled.
Once enabled the "build-in" tools, and ask a request that needs to call the tool, the Gemini model then response with blank output.

It looks like the issue was related with the "built-in" tools with Gemini.

@silenceroom commented on GitHub (Apr 10, 2026): sorry for the late response, super busy in the last few weeks. With multiple rounds of test, I can reproduce the issue with below conditions: 1. Gemini works nomral without "built-in" tools enabled. 2. Once enabled the "build-in" tools, and ask a request that needs to call the tool, the Gemini model then response with blank output. It looks like the issue was related with the "built-in" tools with Gemini.

GiteaMirror commented

2026-05-15 15:52:15 -05:00

@soreikomori commented on GitHub (Apr 21, 2026):

Bug still up on 0.9.1. I should mention that back in the thread when this was first raised (for me at least), @jwhandy3 found a fix, but wasn't able to make a pull request. If anyone can confirm this works in their system, maybe this can be PR'd?

@soreikomori commented on GitHub (Apr 21, 2026): Bug still up on 0.9.1. I should mention that back in the thread when this was first raised (for me at least), @jwhandy3 found a [fix](https://github.com/open-webui/open-webui/discussions/19760#discussioncomment-15274190), but wasn't able to make a pull request. If anyone can confirm this works in their system, maybe this can be PR'd?

Sign in to join this conversation.

Branches Tags

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: github-starred/open-webui#90597