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 #14793] generate API ignores think=false for qwen3.5 (chat API works) #35315

New Issue
Closed
opened 2026-04-22 19:45:20 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-22 19:45:20 -05:00
Owner
Copy Link

Originally created by @andreamoro on GitHub (Mar 12, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14793

What is the issue?

The generate API completely ignores think: false passed in options for qwen3.5:9b. The model still produces thinking tokens that consume the entire num_predict budget, resulting in an empty response field. The chat API with think=False as a top-level parameter works correctly.

This asymmetry causes silent failures for applications using the generate endpoint with thinking models — they get empty responses with no error.

OS / Ollama version

  • OS: Linux 6.17.0-14-generic (Ubuntu)
  • Ollama: 0.17.7
  • Model: qwen3.5:9b

Steps to reproduce

import ollama

# Test 1: generate API — think=false IGNORED
r1 = ollama.generate(
    model='qwen3.5:9b',
    prompt='Say hello in 2 words. Reply with ONLY the words.',
    options={'num_predict': 30, 'think': False}
)
print('generate think=false:')
print('  response:', repr(r1['response']))        # '' (empty!)
print('  thinking:', repr(r1['thinking'][:80]))    # still populated
print('  eval_count:', r1['eval_count'])           # 30 (all tokens burned on thinking)

# Test 2: chat API — think=False WORKS
r2 = ollama.chat(
    model='qwen3.5:9b',
    messages=[{'role': 'user', 'content': 'Say hello in 2 words. Reply with ONLY the words.'}],
    think=False,
    options={'num_predict': 30}
)
print('chat think=False:')
print('  content:', repr(r2['message']['content']))  # 'Hello there!' (actual output)
print('  eval_count:', r2['eval_count'])              # 3 (no wasted tokens)

Actual output

generate think=false:
  response: ''
  thinking: 'Thinking Process:\n\n1.  **Analyze the Request:**\n    *   Task: Say hello.\n    *   Constraint 1:'
  eval_count: 30

chat think=False:
  content: 'Hello there!'
  eval_count: 3

Expected behavior

think: false should disable thinking on BOTH generate and chat endpoints. When thinking is disabled, all num_predict tokens should be available for the visible response.

Additional context

Workaround

Switch from generate to chat API and pass think=False as a top-level parameter (not inside options).

Originally created by @andreamoro on GitHub (Mar 12, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14793 ### What is the issue? The `generate` API completely ignores `think: false` passed in `options` for `qwen3.5:9b`. The model still produces thinking tokens that consume the entire `num_predict` budget, resulting in an empty `response` field. The `chat` API with `think=False` as a top-level parameter works correctly. This asymmetry causes silent failures for applications using the `generate` endpoint with thinking models — they get empty responses with no error. ### OS / Ollama version - OS: Linux 6.17.0-14-generic (Ubuntu) - Ollama: 0.17.7 - Model: qwen3.5:9b ### Steps to reproduce ```python import ollama # Test 1: generate API — think=false IGNORED r1 = ollama.generate( model='qwen3.5:9b', prompt='Say hello in 2 words. Reply with ONLY the words.', options={'num_predict': 30, 'think': False} ) print('generate think=false:') print(' response:', repr(r1['response'])) # '' (empty!) print(' thinking:', repr(r1['thinking'][:80])) # still populated print(' eval_count:', r1['eval_count']) # 30 (all tokens burned on thinking) # Test 2: chat API — think=False WORKS r2 = ollama.chat( model='qwen3.5:9b', messages=[{'role': 'user', 'content': 'Say hello in 2 words. Reply with ONLY the words.'}], think=False, options={'num_predict': 30} ) print('chat think=False:') print(' content:', repr(r2['message']['content'])) # 'Hello there!' (actual output) print(' eval_count:', r2['eval_count']) # 3 (no wasted tokens) ``` ### Actual output ``` generate think=false: response: '' thinking: 'Thinking Process:\n\n1. **Analyze the Request:**\n * Task: Say hello.\n * Constraint 1:' eval_count: 30 chat think=False: content: 'Hello there!' eval_count: 3 ``` ### Expected behavior `think: false` should disable thinking on BOTH `generate` and `chat` endpoints. When thinking is disabled, all `num_predict` tokens should be available for the visible response. ### Additional context - Non-thinking models (e.g. `gemma2:9b`) are unaffected — `think: false` is harmlessly ignored - The model's Modelfile uses `RENDERER qwen3.5` / `PARSER qwen3.5`, so thinking is handled by Ollama's built-in renderer, not the template - Increasing `num_predict` to 200+ doesn't help on `generate` — the model just thinks longer, still empty `response` - Related: #14502, #14612, #14645 ### Workaround Switch from `generate` to `chat` API and pass `think=False` as a top-level parameter (not inside `options`).
GiteaMirror added the thinking label 2026-04-22 19:45:20 -05:00
GiteaMirror commented 2026-04-22 19:45:22 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 12, 2026):

--- 14793.py.orig	2026-03-12 10:35:55.053215178 +0100
+++ 14793.py	2026-03-12 10:41:33.575422728 +0100
@@ -4,11 +4,12 @@
 r1 = ollama.generate(
     model='qwen3.5:9b',
     prompt='Say hello in 2 words. Reply with ONLY the words.',
-    options={'num_predict': 30, 'think': False}
+    think=False,
+    options={'num_predict': 30}
 )
 print('generate think=false:')
 print('  response:', repr(r1['response']))        # '' (empty!)
-print('  thinking:', repr(r1['thinking'][:80]))    # still populated
+print('  thinking:', repr(r1.get('thinking')))    # still populated
 print('  eval_count:', r1['eval_count'])           # 30 (all tokens burned on thinking)
 
 # Test 2: chat API — think=False WORKS
<!-- gh-comment-id:4045318400 --> @rick-github commented on GitHub (Mar 12, 2026): ```diff --- 14793.py.orig 2026-03-12 10:35:55.053215178 +0100 +++ 14793.py 2026-03-12 10:41:33.575422728 +0100 @@ -4,11 +4,12 @@ r1 = ollama.generate( model='qwen3.5:9b', prompt='Say hello in 2 words. Reply with ONLY the words.', - options={'num_predict': 30, 'think': False} + think=False, + options={'num_predict': 30} ) print('generate think=false:') print(' response:', repr(r1['response'])) # '' (empty!) -print(' thinking:', repr(r1['thinking'][:80])) # still populated +print(' thinking:', repr(r1.get('thinking'))) # still populated print(' eval_count:', r1['eval_count']) # 30 (all tokens burned on thinking) # Test 2: chat API — think=False WORKS ```
GiteaMirror commented 2026-04-22 19:45:23 -05:00
Author
Owner
Copy Link

@andreamoro commented on GitHub (Mar 12, 2026):

Thanks @rick-github — confirmed! think=False as a top-level parameter works correctly on both generate and chat.

The actual issue is that think inside options={} is silently ignored. Updated test:

import ollama

# BROKEN: think inside options — silently ignored, empty response
r1 = ollama.generate(
    model='qwen3.5:9b',
    prompt='Say hello in 2 words.',
    options={'num_predict': 30, 'think': False}
)
print(r1['response'])  # '' (empty!)

# WORKS: think as top-level parameter
r2 = ollama.generate(
    model='qwen3.5:9b',
    prompt='Say hello in 2 words.',
    think=False,
    options={'num_predict': 30}
)
print(r2['response'])  # 'Hello there'

This is a silent failure that's easy to hit — think looks like a model parameter (similar to temperature, num_predict) so placing it in options feels natural. Perhaps Ollama could warn when unrecognized keys appear in options?

<!-- gh-comment-id:4046561008 --> @andreamoro commented on GitHub (Mar 12, 2026): Thanks @rick-github — confirmed! `think=False` as a top-level parameter works correctly on both `generate` and `chat`. The actual issue is that `think` inside `options={}` is silently ignored. Updated test: ```python import ollama # BROKEN: think inside options — silently ignored, empty response r1 = ollama.generate( model='qwen3.5:9b', prompt='Say hello in 2 words.', options={'num_predict': 30, 'think': False} ) print(r1['response']) # '' (empty!) # WORKS: think as top-level parameter r2 = ollama.generate( model='qwen3.5:9b', prompt='Say hello in 2 words.', think=False, options={'num_predict': 30} ) print(r2['response']) # 'Hello there' ``` This is a silent failure that's easy to hit — `think` looks like a model parameter (similar to `temperature`, `num_predict`) so placing it in `options` feels natural. Perhaps Ollama could warn when unrecognized keys appear in `options`?
GiteaMirror commented 2026-04-22 19:45:23 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 12, 2026):

time=2026-03-12T14:03:45.429Z level=WARN source=types.go:976 msg="invalid option provided" option=think
<!-- gh-comment-id:4047042924 --> @rick-github commented on GitHub (Mar 12, 2026): ``` time=2026-03-12T14:03:45.429Z level=WARN source=types.go:976 msg="invalid option provided" option=think ```
GiteaMirror commented 2026-04-22 19:45:24 -05:00
Author
Owner
Copy Link

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

I'm still running into this issue; why not allow us to manage the thinking parameter ourselves?

<!-- gh-comment-id:4146770722 --> @duaneking commented on GitHub (Mar 28, 2026): I'm still running into this issue; why not allow us to manage the thinking parameter ourselves?
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 thinking
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#35315
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?