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 #8529] feat: OpenAI reasoning_content compatibility #67556

New Issue
Closed
opened 2026-05-04 10:48:29 -05:00 by GiteaMirror · 15 comments
GiteaMirror commented 2026-05-04 10:48:29 -05:00
Owner
Copy Link

Originally created by @EntropyYue on GitHub (Jan 22, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/8529

Originally assigned to: @drifkin on GitHub.

The current thinking model outputs XML tags to distinguish between thinking and answering, and we need the reasoning_content feature

Originally created by @EntropyYue on GitHub (Jan 22, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/8529 Originally assigned to: @drifkin on GitHub. The current thinking model outputs XML tags to distinguish between thinking and answering, and we need the `reasoning_content` feature
GiteaMirror added the feature request label 2026-05-04 10:48:29 -05:00
GiteaMirror commented 2026-05-04 10:48:34 -05:00
Author
Owner
Copy Link

@sunburst-yz commented on GitHub (Jan 22, 2025):

+1

<!-- gh-comment-id:2606470064 --> @sunburst-yz commented on GitHub (Jan 22, 2025): +1
GiteaMirror commented 2026-05-04 10:48:36 -05:00
Author
Owner
Copy Link

@whisper-bye commented on GitHub (Feb 3, 2025):

should support

reasoning_content = response.choices[0].message.reasoning_content
content = response.choices[0].message.content
<!-- gh-comment-id:2631829987 --> @whisper-bye commented on GitHub (Feb 3, 2025): should support ``` reasoning_content = response.choices[0].message.reasoning_content content = response.choices[0].message.content ```
GiteaMirror commented 2026-05-04 10:48:38 -05:00
Author
Owner
Copy Link

@goactiongo commented on GitHub (Feb 6, 2025):

+1

<!-- gh-comment-id:2638967449 --> @goactiongo commented on GitHub (Feb 6, 2025): +1
GiteaMirror commented 2026-05-04 10:48:40 -05:00
Author
Owner
Copy Link

@leslie2046 commented on GitHub (Feb 10, 2025):

+1

<!-- gh-comment-id:2646791167 --> @leslie2046 commented on GitHub (Feb 10, 2025): +1
GiteaMirror commented 2026-05-04 10:48:43 -05:00
Author
Owner
Copy Link

@asgraf commented on GitHub (Feb 10, 2025):

+1

<!-- gh-comment-id:2646991526 --> @asgraf commented on GitHub (Feb 10, 2025): +1
GiteaMirror commented 2026-05-04 10:48:45 -05:00
Author
Owner
Copy Link

@kasnerz commented on GitHub (Feb 12, 2025):

Separating these two would also enable reasoning along structured outputs, right? Or does that need to be implemented separately?

For clarification: with deepseek-r1, structured outputs are currently working but I assume it eliminates the reasoning trace.

<!-- gh-comment-id:2653717646 --> @kasnerz commented on GitHub (Feb 12, 2025): Separating these two would also enable reasoning along structured outputs, right? Or does that need to be implemented separately? For clarification: with deepseek-r1, structured outputs are currently working but I assume it eliminates the reasoning trace.
GiteaMirror commented 2026-05-04 10:48:47 -05:00
Author
Owner
Copy Link

@wuyanfeiwork commented on GitHub (Feb 15, 2025):

+1

<!-- gh-comment-id:2660960423 --> @wuyanfeiwork commented on GitHub (Feb 15, 2025): +1
GiteaMirror commented 2026-05-04 10:48:48 -05:00
Author
Owner
Copy Link

@danny-avila commented on GitHub (Feb 15, 2025):

+1

This should be the standard over the XML tags to avoid parsing text for reasoning tokens.

<!-- gh-comment-id:2661046439 --> @danny-avila commented on GitHub (Feb 15, 2025): +1 This should be the standard over the XML tags to avoid parsing text for reasoning tokens.
GiteaMirror commented 2026-05-04 10:48:50 -05:00
Author
Owner
Copy Link

@gongzhang commented on GitHub (Feb 19, 2025):

+1

The <think> and </think> are special tokens in the tokenizer and should not be directly output as plain string. With the current use of XML tags, it cannot properly handle prompts like "Please output ten <think> and ten </think>, one per line."

<!-- gh-comment-id:2667974605 --> @gongzhang commented on GitHub (Feb 19, 2025): +1 The `<think>` and `</think>` are special tokens in the tokenizer and **should not be directly output as plain string**. With the current use of XML tags, it cannot properly handle prompts like "Please output ten \<think> and ten <\/think>, one per line."
GiteaMirror commented 2026-05-04 10:48:51 -05:00
Author
Owner
Copy Link

@wuyanfeiwork commented on GitHub (Feb 19, 2025):

I trust that the team will address and optimize this issue promptly. #9137

<!-- gh-comment-id:2668080722 --> @wuyanfeiwork commented on GitHub (Feb 19, 2025): I trust that the team will address and optimize this issue promptly. #9137
GiteaMirror commented 2026-05-04 10:48:53 -05:00
Author
Owner
Copy Link

@anunknowperson commented on GitHub (May 2, 2025):

This issue has become much more prevalent with the release of the Qwen3 models. With the correct prompt template from transformers, an empty \n block is always added to the assistant's response if thought mode is turned off. It would be more convenient if ollama could parse it automatically into reasoning content. With thought mode enabled (supposedly there should be a prefill, which ollama doesn't do in its prompt template either), it would be better if thoughts were in reasoning_content (though reasoning parsing must be a feature with ability to enable/disable).

<!-- gh-comment-id:2847074997 --> @anunknowperson commented on GitHub (May 2, 2025): This issue has become much more prevalent with the release of the Qwen3 models. With the correct prompt template from transformers, an empty <think>\n</think> block is always added to the assistant's response if thought mode is turned off. It would be more convenient if ollama could parse it automatically into reasoning content. With thought mode enabled (supposedly there should be a prefill, which ollama doesn't do in its prompt template either), it would be better if thoughts were in reasoning_content (though reasoning parsing must be a feature with ability to enable/disable).
GiteaMirror commented 2026-05-04 10:48:54 -05:00
Author
Owner
Copy Link

@ParthSareen commented on GitHub (May 16, 2025):

We're getting thinking separation in the Ollama API but I don't see any reasoning being provided in the OpenAI API for compatibility: https://platform.openai.com/docs/guides/reasoning?api-mode=responses

I might be looking at the wrong doc feel free to link the right one.

<!-- gh-comment-id:2887619397 --> @ParthSareen commented on GitHub (May 16, 2025): We're getting thinking separation in the Ollama API but I don't see any reasoning being provided in the OpenAI API for compatibility: https://platform.openai.com/docs/guides/reasoning?api-mode=responses I might be looking at the wrong doc feel free to link the right one.
GiteaMirror commented 2026-05-04 10:48:55 -05:00
Author
Owner
Copy Link

@williamlzw commented on GitHub (May 20, 2025):

+1
https://github.com/imxcstar/Tinvo/blob/master/Tinvo.Provider.OpenAI/AIScheduler/OpenAIProviderParser.cs
https://api-docs.deepseek.com/guides/reasoning_model
This is the reasoning_content output, and currently deepseek-r1 and qwen3 support this output.

from openai import OpenAI
client = OpenAI(api_key="<DeepSeek API Key>", base_url="https://api.deepseek.com")

# Round 1
messages = [{"role": "user", "content": "9.11 and 9.8, which is greater?"}]
response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=messages,
    stream=True
)

reasoning_content = ""
content = ""

for chunk in response:
    if chunk.choices[0].delta.reasoning_content:
        reasoning_content += chunk.choices[0].delta.reasoning_content
    else:
        content += chunk.choices[0].delta.content

# Round 2
messages.append({"role": "assistant", "content": content})
messages.append({'role': 'user', 'content': "How many Rs are there in the word 'strawberry'?"})
response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=messages,
    stream=True
)
<!-- gh-comment-id:2893161174 --> @williamlzw commented on GitHub (May 20, 2025): +1 https://github.com/imxcstar/Tinvo/blob/master/Tinvo.Provider.OpenAI/AIScheduler/OpenAIProviderParser.cs https://api-docs.deepseek.com/guides/reasoning_model This is the reasoning_content output, and currently deepseek-r1 and qwen3 support this output. ``` from openai import OpenAI client = OpenAI(api_key="<DeepSeek API Key>", base_url="https://api.deepseek.com") # Round 1 messages = [{"role": "user", "content": "9.11 and 9.8, which is greater?"}] response = client.chat.completions.create( model="deepseek-reasoner", messages=messages, stream=True ) reasoning_content = "" content = "" for chunk in response: if chunk.choices[0].delta.reasoning_content: reasoning_content += chunk.choices[0].delta.reasoning_content else: content += chunk.choices[0].delta.content # Round 2 messages.append({"role": "assistant", "content": content}) messages.append({'role': 'user', 'content': "How many Rs are there in the word 'strawberry'?"}) response = client.chat.completions.create( model="deepseek-reasoner", messages=messages, stream=True ) ```
GiteaMirror commented 2026-05-04 10:48:58 -05:00
Author
Owner
Copy Link

@danny-avila commented on GitHub (May 25, 2025):

Does PR #10584 address programmatic differentiation between reasoning and final output?

My main need, as I suspect others', is distinguishing "thinking" from answers without text parsing - ideally following Deepseek-R1's standard with separate fields, but any structured approach would work.

<!-- gh-comment-id:2907858055 --> @danny-avila commented on GitHub (May 25, 2025): Does PR #10584 address programmatic differentiation between reasoning and final output? My main need, as I suspect others', is distinguishing "thinking" from answers without text parsing - ideally following Deepseek-R1's standard with separate fields, but any structured approach would work.
GiteaMirror commented 2026-05-04 10:49:02 -05:00
Author
Owner
Copy Link

@ParthSareen commented on GitHub (May 27, 2025):

Yes @danny-avila, there are separate fields to identify and enable thinking content as well as get a final output returned.

<!-- gh-comment-id:2910893497 --> @ParthSareen commented on GitHub (May 27, 2025): Yes @danny-avila, there are separate fields to identify and enable thinking content as well as get a final output returned.
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 feature request
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#67556
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?