github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14567] Bug: gemini-3-flash-preview:cloud - Function Calling returns 400 (missing thought_signature) #71506

New Issue
Open
opened 2026-05-05 01:57:29 -05:00 by GiteaMirror · 10 comments
GiteaMirror commented 2026-05-05 01:57:29 -05:00
Owner
Copy Link

Originally created by @rovanni on GitHub (Mar 3, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14567

What is the issue?

Description

When using gemini-3-flash-preview:cloud via Ollama Cloud with function calling / tool calling enabled, the API returns a 400 Bad Request error whenever the model triggers a tool call:

Ollama API error 400: {"StatusCode":400,"Status":"400 Bad Request","error":"Function call is missing a thought_signature in functionCall parts."}

This indicates that Ollama's function-calling integration for Gemini 3 does not yet handle the thought_signature field required by Google's Gemini 3 API for tool calls.

Steps to Reproduce

  1. Configure Ollama Cloud with the gemini-3-flash-preview:cloud model.
  2. Define one or more tools/functions in the Ollama request payload.
  3. Send a prompt that causes the model to invoke a tool.
  4. Observe the HTTP 400 error with the message above.

Expected Behavior

Tool / function calling with gemini-3-flash-preview:cloud should work the same way as other Ollama Cloud models (e.g. glm-5:cloud, kimi-k2.5:cloud) — without requiring the client to manually manage or inject a thought_signature field.

Ollama should transparently handle any Gemini 3–specific protocol details (such as thought signatures) when forwarding tool calls and tool results.

Actual Behavior

When the model emits a function/tool call, Ollama forwards the call to the Gemini 3 API but the follow-up request fails with:

Function call is missing a thought_signature in functionCall parts.

According to Google's Gemini 3 documentation, function calls include a thought_signature that must be preserved and echoed back with tool results. Ollama does not currently handle this field, causing all tool-calling flows to fail for this model.

Environment

Field Value
Ollama Version 0.17.4
Model (failing) gemini-3-flash-preview:cloud
Models (working) glm-5:cloud, kimi-k2.5:cloud
Error Function call is missing a thought_signature in functionCall parts

Additional Context

  • Google's Gemini 3 API enforces a thought_signature mechanism: every function call response includes a signature that must be sent back alongside the tool result; omitting it causes 400 errors.
  • Multiple SDKs integrating Gemini 3 have had to add explicit support for thought_signature to make tool calling work.
  • Ollama Cloud currently does not propagate or manage thought_signature for Gemini 3, causing tool integrations that work with other models to fail specifically for gemini-3-flash-preview:cloud.

Questions:

  • Is support for Gemini 3 thought_signature in tool/function calling already planned?
  • Is there a recommended workaround in the meantime (e.g. disable tools for Gemini 3, or use a different Gemini model)?

Relevant log output

Ollama API error 400: {"StatusCode":400,"Status":"400 Bad Request","error":"Function call is missing a thought_signature in functionCall parts."}

OS

Linux

GPU

No response

CPU

No response

Ollama version

0.17.4

Originally created by @rovanni on GitHub (Mar 3, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14567 ### What is the issue? ## Description When using `gemini-3-flash-preview:cloud` via Ollama Cloud with function calling / tool calling enabled, the API returns a **400 Bad Request** error whenever the model triggers a tool call: ``` Ollama API error 400: {"StatusCode":400,"Status":"400 Bad Request","error":"Function call is missing a thought_signature in functionCall parts."} ``` This indicates that Ollama's function-calling integration for Gemini 3 does not yet handle the `thought_signature` field required by Google's Gemini 3 API for tool calls. --- ## Steps to Reproduce 1. Configure Ollama Cloud with the `gemini-3-flash-preview:cloud` model. 2. Define one or more tools/functions in the Ollama request payload. 3. Send a prompt that causes the model to invoke a tool. 4. Observe the HTTP 400 error with the message above. --- ## Expected Behavior Tool / function calling with `gemini-3-flash-preview:cloud` should work the same way as other Ollama Cloud models (e.g. `glm-5:cloud`, `kimi-k2.5:cloud`) — without requiring the client to manually manage or inject a `thought_signature` field. Ollama should transparently handle any Gemini 3–specific protocol details (such as thought signatures) when forwarding tool calls and tool results. --- ## Actual Behavior When the model emits a function/tool call, Ollama forwards the call to the Gemini 3 API but the follow-up request fails with: ``` Function call is missing a thought_signature in functionCall parts. ``` According to Google's Gemini 3 documentation, function calls include a `thought_signature` that must be preserved and echoed back with tool results. Ollama does not currently handle this field, causing all tool-calling flows to fail for this model. --- ## Environment | Field | Value | |---------------------|--------------------------------------------------------------------| | Ollama Version | 0.17.4 | | Model (failing) | `gemini-3-flash-preview:cloud` | | Models (working) | `glm-5:cloud`, `kimi-k2.5:cloud` | | Error | `Function call is missing a thought_signature in functionCall parts` | --- ## Additional Context - Google's Gemini 3 API enforces a `thought_signature` mechanism: every function call response includes a signature that must be sent back alongside the tool result; omitting it causes 400 errors. - Multiple SDKs integrating Gemini 3 have had to add explicit support for `thought_signature` to make tool calling work. - Ollama Cloud currently does not propagate or manage `thought_signature` for Gemini 3, causing tool integrations that work with other models to fail specifically for `gemini-3-flash-preview:cloud`. **Questions:** - Is support for Gemini 3 `thought_signature` in tool/function calling already planned? - Is there a recommended workaround in the meantime (e.g. disable tools for Gemini 3, or use a different Gemini model)? ### Relevant log output ```shell Ollama API error 400: {"StatusCode":400,"Status":"400 Bad Request","error":"Function call is missing a thought_signature in functionCall parts."} ``` ### OS Linux ### GPU _No response_ ### CPU _No response_ ### Ollama version 0.17.4
GiteaMirror added the bug label 2026-05-05 01:57:29 -05:00
GiteaMirror commented 2026-05-05 01:57:30 -05:00
Author
Owner
Copy Link

@willificent commented on GitHub (Mar 4, 2026):

Additional Finding: Tool Call Arguments Format Mismatch

I've been debugging this same issue and discovered a related problem that may help isolate the bug.

The Arguments Format Issue

When testing gemini-3-flash-preview:cloud directly via /api/chat, I found that:

  1. Gemini returns tool call arguments as an object, not a string:
{
  "index": 0,
  "name": "get_weather",
  "arguments": {
    "city": "Chicago"
  }
}
  1. Sending tool results with arguments as a JSON string fails with a different error:
{"error":"Value looks like object, but can't find closing '}' symbol"}
  1. Sending tool results with arguments as an object works — the round-trip succeeds when the arguments are passed as a proper object rather than a stringified JSON.

Test Cases

Fails (arguments as JSON string — OpenAI-compatible format):

{"id": "tc1", "function": {"name": "get_weather", "arguments": "{\"city\": \"Chicago\"}"}}

Works (arguments as object):

{"id": "tc1", "function": {"name": "get_weather", "arguments": {"city": "Chicago"}}}

Hypothesis

It appears there are two separate issues affecting Gemini 3 tool calling:

  1. Arguments format mismatch — Ollama's OpenAI-compatible layer may be expecting stringified arguments, but Gemini Cloud returns objects. This causes the parsing error above.

  2. Missing thought_signature — The original issue reported. Even when the format is correct, Gemini 3 requires a thought_signature field that Ollama doesn't currently preserve/forward.

The first issue can be worked around by ensuring arguments are passed as objects, but the second is a protocol-level gap that requires Ollama to track and inject the signature from the initial tool call response.

Environment

Field Value
Ollama Version 0.17.4
Model gemini-3-flash-preview:cloud
Endpoint /api/chat (native Ollama API)

Hope this helps narrow down the fix!

<!-- gh-comment-id:4000240770 --> @willificent commented on GitHub (Mar 4, 2026): ## Additional Finding: Tool Call Arguments Format Mismatch I've been debugging this same issue and discovered a related problem that may help isolate the bug. ### The Arguments Format Issue When testing `gemini-3-flash-preview:cloud` directly via `/api/chat`, I found that: 1. **Gemini returns tool call arguments as an object**, not a string: ```json { "index": 0, "name": "get_weather", "arguments": { "city": "Chicago" } } ``` 2. **Sending tool results with arguments as a JSON string fails** with a different error: ``` {"error":"Value looks like object, but can't find closing '}' symbol"} ``` 3. **Sending tool results with arguments as an object works** — the round-trip succeeds when the arguments are passed as a proper object rather than a stringified JSON. ### Test Cases **Fails** (arguments as JSON string — OpenAI-compatible format): ```json {"id": "tc1", "function": {"name": "get_weather", "arguments": "{\"city\": \"Chicago\"}"}} ``` **Works** (arguments as object): ```json {"id": "tc1", "function": {"name": "get_weather", "arguments": {"city": "Chicago"}}} ``` ### Hypothesis It appears there are **two separate issues** affecting Gemini 3 tool calling: 1. **Arguments format mismatch** — Ollama's OpenAI-compatible layer may be expecting stringified arguments, but Gemini Cloud returns objects. This causes the parsing error above. 2. **Missing `thought_signature`** — The original issue reported. Even when the format is correct, Gemini 3 requires a `thought_signature` field that Ollama doesn't currently preserve/forward. The first issue can be worked around by ensuring arguments are passed as objects, but the second is a protocol-level gap that requires Ollama to track and inject the signature from the initial tool call response. ### Environment | Field | Value | |-------|-------| | Ollama Version | 0.17.4 | | Model | `gemini-3-flash-preview:cloud` | | Endpoint | `/api/chat` (native Ollama API) | Hope this helps narrow down the fix!
GiteaMirror commented 2026-05-05 01:57:31 -05:00
Author
Owner
Copy Link

@chereekana-droid commented on GitHub (Mar 6, 2026):

Hopefully, this issue will be taken seriously; it's just a simple bug.

<!-- gh-comment-id:4010512473 --> @chereekana-droid commented on GitHub (Mar 6, 2026): Hopefully, this issue will be taken seriously; it's just a simple bug.
GiteaMirror commented 2026-05-05 01:57:31 -05:00
Author
Owner
Copy Link

@willificent commented on GitHub (Mar 6, 2026):

Any progress on this perhaps? It is a super simple bug and, compared to GLM-5, Gemini 3 Flash is almost exactly as smart and nearly 3x as fast. I'll see if I can do some work on a fix and submit a PR.

<!-- gh-comment-id:4013117952 --> @willificent commented on GitHub (Mar 6, 2026): Any progress on this perhaps? It is a super simple bug and, compared to GLM-5, Gemini 3 Flash is almost exactly as smart and nearly 3x as fast. I'll see if I can do some work on a fix and submit a PR.
GiteaMirror commented 2026-05-05 01:57:32 -05:00
Author
Owner
Copy Link

@Tattooed-Geek commented on GitHub (Mar 6, 2026):

I am also waiting for a fix! 💪😜

<!-- gh-comment-id:4013135848 --> @Tattooed-Geek commented on GitHub (Mar 6, 2026): I am also waiting for a fix! 💪😜
GiteaMirror commented 2026-05-05 01:57:33 -05:00
Author
Owner
Copy Link

@willificent commented on GitHub (Mar 6, 2026):

Okay, I vibe coded a fix, fingers crossed this checks out at validation LOL https://github.com/ollama/ollama/pull/14676

<!-- gh-comment-id:4014001179 --> @willificent commented on GitHub (Mar 6, 2026): Okay, I vibe coded a fix, fingers crossed this checks out at validation LOL https://github.com/ollama/ollama/pull/14676
GiteaMirror commented 2026-05-05 01:57:33 -05:00
Author
Owner
Copy Link

@reflexmrl commented on GitHub (Mar 11, 2026):

Confirming this issue with OpenClaw + Gemini 3 via Ollama Cloud. Exact same error.
PR #14676 looks like the fix — looking forward to the merge! 🦞

<!-- gh-comment-id:4041478038 --> @reflexmrl commented on GitHub (Mar 11, 2026): Confirming this issue with OpenClaw + Gemini 3 via Ollama Cloud. Exact same error. PR #14676 looks like the fix — looking forward to the merge! 🦞
GiteaMirror commented 2026-05-05 01:57:33 -05:00
Author
Owner
Copy Link

@chereekana-droid commented on GitHub (Mar 14, 2026):

What's the latest development, my friends?

<!-- gh-comment-id:4059184598 --> @chereekana-droid commented on GitHub (Mar 14, 2026): What's the latest development, my friends?
GiteaMirror commented 2026-05-05 01:57:34 -05:00
Author
Owner
Copy Link

@sakuradonut commented on GitHub (Apr 14, 2026):

看起來依然沒有新的進展，gema4的修復都做了，但gemini的這個400問題ollama依舊擱置。

<!-- gh-comment-id:4241409963 --> @sakuradonut commented on GitHub (Apr 14, 2026): 看起來依然沒有新的進展，gema4的修復都做了，但gemini的這個400問題ollama依舊擱置。
GiteaMirror commented 2026-05-05 01:57:34 -05:00
Author
Owner
Copy Link

@willificent commented on GitHub (Apr 16, 2026):

说起来，Ollama 还专门发了新版来支持 Gemma4，但 Gemini 3 Flash 好像没什么人关心?

<!-- gh-comment-id:4261126591 --> @willificent commented on GitHub (Apr 16, 2026): 说起来，Ollama 还专门发了新版来支持 Gemma4，但 Gemini 3 Flash 好像没什么人关心?
GiteaMirror commented 2026-05-05 01:57:35 -05:00
Author
Owner
Copy Link

@krauserene89 commented on GitHub (Apr 28, 2026):

Hi team, any ETA on when this might be fixed? Would love to use Gemini 3 Flash with tool calling — it's a great model for my use case. Thanks for the awesome work! 🙏

<!-- gh-comment-id:4333284480 --> @krauserene89 commented on GitHub (Apr 28, 2026): Hi team, any ETA on when this might be fixed? Would love to use Gemini 3 Flash with tool calling — it's a great model for my use case. Thanks for the awesome work! 🙏
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#71506
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?