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 #15293] gemma4 thinking not forwarded via /v1/chat/completions #71845

New Issue
Open
opened 2026-05-05 02:41:29 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-05-05 02:41:29 -05:00
Owner
Copy Link

Originally created by @roamiiing on GitHub (Apr 3, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15293

Description

When using Gemma 4 models, the think: true parameter is ignored by the
/v1/chat/completions endpoint. The same parameter works correctly via
the native /api/chat endpoint.

Steps to reproduce

Native API (works ✓):
curl http://127.0.0.1:11434/api/chat
-d '{"model":"gemma4:e4b","think":true,"stream":true, "messages":[{"role":"user","content":"What is 2+2?"}]}'

Response includes "thinking" field with reasoning content

OpenAI-compatible API (broken ✗):
curl http://127.0.0.1:11434/v1/chat/completions
-H "Content-Type: application/json"
-d '{"model":"gemma4:e4b","think":true,"stream":true, "messages":[{"role":"user","content":"What is 2+2?"}]}'

Response: just "4", no thinking content
Also tried: "options":{"think":true}, "reasoning_effort":"high" — none work.

Environment

  • Ollama version: 0.20.0
  • Model: gemma4:e4b
  • OS: Linux (WSL2)

Expected behavior

think:true should be forwarded to the model when using /v1/,
same as it was fixed for deepseek-r1 in #15036.

Originally created by @roamiiing on GitHub (Apr 3, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15293 ## Description When using Gemma 4 models, the `think: true` parameter is ignored by the `/v1/chat/completions` endpoint. The same parameter works correctly via the native `/api/chat` endpoint. ## Steps to reproduce Native API (works ✓): curl http://127.0.0.1:11434/api/chat -d '{"model":"gemma4:e4b","think":true,"stream":true, "messages":[{"role":"user","content":"What is 2+2?"}]}' Response includes "thinking" field with reasoning content OpenAI-compatible API (broken ✗): curl http://127.0.0.1:11434/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"gemma4:e4b","think":true,"stream":true, "messages":[{"role":"user","content":"What is 2+2?"}]}' Response: just "4", no thinking content Also tried: `"options":{"think":true}`, `"reasoning_effort":"high"` — none work. ## Environment - Ollama version: 0.20.0 - Model: gemma4:e4b - OS: Linux (WSL2) ## Expected behavior `think:true` should be forwarded to the model when using `/v1/`, same as it was fixed for deepseek-r1 in #15036.
GiteaMirror commented 2026-05-05 02:41:30 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 3, 2026):

Use reasoning_effort and give it a prompt it has to think about.

$ curl -s http://127.0.0.1:11434/v1/chat/completions -d '{
  "model":"gemma4:e4b",
  "reasoning_effort":"low",
  "messages":[{"role":"user","content":"Why is the sky blue?"}]
}' | jq -r '.choices[0].message|"reasoning: \(.reasoning[0:50])\ncontent  : \(.content[0:50])"'
reasoning: Here's a thinking process that leads to the sugges
content  : This is one of the most beautiful and common quest
<!-- gh-comment-id:4184159137 --> @rick-github commented on GitHub (Apr 3, 2026): Use `reasoning_effort` and give it a prompt it has to think about. ```console $ curl -s http://127.0.0.1:11434/v1/chat/completions -d '{ "model":"gemma4:e4b", "reasoning_effort":"low", "messages":[{"role":"user","content":"Why is the sky blue?"}] }' | jq -r '.choices[0].message|"reasoning: \(.reasoning[0:50])\ncontent : \(.content[0:50])"' reasoning: Here's a thinking process that leads to the sugges content : This is one of the most beautiful and common quest ```
GiteaMirror commented 2026-05-05 02:41:30 -05:00
Author
Owner
Copy Link

@verdverm commented on GitHub (Apr 7, 2026):

The gemma 4 model card mentions multiple thinking levels, how do those map onto the available options in the openai compat api?

This seems like a general question or documentation gap for most thinking models.

For gemma4, it looks like all mentioned levels [low,medium,high] look to trigger thoughts, however "none" does as well, which seems like a bug. It's unclear if the different values have meaning to gemma

<!-- gh-comment-id:4196791210 --> @verdverm commented on GitHub (Apr 7, 2026): The gemma 4 model card mentions multiple thinking levels, how do those map onto the available options in the openai compat api? This seems like a general question or documentation gap for most thinking models. For gemma4, it looks like all mentioned levels [low,medium,high] look to trigger thoughts, however "none" does as well, which seems like a bug. It's unclear if the different values have meaning to gemma
GiteaMirror commented 2026-05-05 02:41:31 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 7, 2026):

The original template for the model doesn't indicate support for multiple thinking levels. Perhaps "configurable thinking modes" from the model card means "on" and "off".

$ for t in high medium low none ; do r="$(curl -s localhost:11434/v1/chat/completions -d '{"model":"gemma4","messages":[{"role":"user","content":"why is the sky blue"}],"reas
oning_effort":"'$t'","stream":false,"seed":0,"temperature":0}' | jq -r '.choices[0].message.reasoning//""')" ; printf "%-7s %4d %s\n" $t ${#r} "${r::20}" ; done
high    2607 Here's a thinking pr
medium  2607 Here's a thinking pr
low     2607 Here's a thinking pr
none       0
<!-- gh-comment-id:4198643909 --> @rick-github commented on GitHub (Apr 7, 2026): The [original template](https://huggingface.co/google/gemma-4-E2B-it/blob/main/chat_template.jinja) for the model doesn't indicate support for multiple thinking levels. Perhaps "configurable thinking modes" from the model card means "on" and "off". ```console $ for t in high medium low none ; do r="$(curl -s localhost:11434/v1/chat/completions -d '{"model":"gemma4","messages":[{"role":"user","content":"why is the sky blue"}],"reas oning_effort":"'$t'","stream":false,"seed":0,"temperature":0}' | jq -r '.choices[0].message.reasoning//""')" ; printf "%-7s %4d %s\n" $t ${#r} "${r::20}" ; done high 2607 Here's a thinking pr medium 2607 Here's a thinking pr low 2607 Here's a thinking pr none 0 ```
GiteaMirror commented 2026-05-05 02:41:31 -05:00
Author
Owner
Copy Link

@verdverm commented on GitHub (Apr 7, 2026):

perhaps, wish they made it less ambiguous

based on https://ai.google.dev/gemini-api/docs/thinking#thinking-levels, Gemini has distinct levels and Gemma is made from Gemini

Is it possible their template is wrong?

<!-- gh-comment-id:4200325593 --> @verdverm commented on GitHub (Apr 7, 2026): perhaps, wish they made it less ambiguous based on https://ai.google.dev/gemini-api/docs/thinking#thinking-levels, Gemini has distinct levels and Gemma is made from Gemini Is it possible their template is wrong?
GiteaMirror commented 2026-05-05 02:41:32 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 7, 2026):

Is it possible their template is wrong?

Pretty sure Google would know how to write a template for their model. According to the documentation, it only supports enabling and disabling.

<!-- gh-comment-id:4200379211 --> @rick-github commented on GitHub (Apr 7, 2026): > Is it possible their template is wrong? Pretty sure Google would know how to write a template for their model. According to the [documentation](https://huggingface.co/google/gemma-4-E4B-it#2-thinking-mode-configuration), it only supports enabling and disabling.
GiteaMirror commented 2026-05-05 02:41:35 -05:00
Author
Owner
Copy Link

@verdverm commented on GitHub (Apr 7, 2026):

thanks for the clarity @rick-github, I learned a few things along the way!

<!-- gh-comment-id:4200827696 --> @verdverm commented on GitHub (Apr 7, 2026): thanks for the clarity @rick-github, I learned a few things along the way!
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
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#71845
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?