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 #13353] Qwen3VLRenderer/Parser ignores think API parameter #70877

New Issue
Closed
opened 2026-05-04 23:19:38 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-05-04 23:19:38 -05:00
Owner
Copy Link

Originally created by @yuu104 on GitHub (Dec 6, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/13353

What is the issue?

When using qwen3-vl model with RENDERER qwen3-vl-thinking (which is the default for the official qwen3-vl model), the think API parameter is completely ignored.

Expected behavior:

  • think: true → thinking mode enabled
  • think: false → thinking mode disabled
  • think not specified → thinking mode disabled (consistent with CogitoRenderer)

Actual behavior: The model always operates in thinking mode regardless of the think parameter, resulting in slow responses (50+ seconds).

API request example

curl http://localhost:11434/api/chat -d '{
  "model": "qwen3-vl",
  "messages": [{"role": "user", "content": "Hello"}],
  "think": false,
  "stream": false
}'

Even with think: false, the model still enters thinking mode.

Root cause (code analysis)

1. Renderer issue

model/renderers/qwen3vl.go:74:

func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) {

The thinkValue parameter is discarded with _. The renderer only uses r.isThinking which is set at construction time based on the renderer name (qwen3-vl-thinking vs qwen3-vl-instruct).

2. Parser issue

model/parsers/qwen3vl.go:57-61:

func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
    p.tools = tools
    p.setInitialState(lastMessage)  // thinkValue is NOT passed here
    return tools
}

thinkValue is received but not passed to setInitialState, so it has no effect.

Comparison with working implementation

CogitoRenderer correctly respects the think parameter (model/renderers/cogito.go:14-20):

func (r *CogitoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
    // ...
    // thinking is enabled: model must support it AND user must request it (true)
    enableThinking := r.isThinking && (thinkValue != nil && thinkValue.Bool())
    // ...
}

This pattern should be applied to Qwen3VLRenderer and Qwen3VLParser as well.

Proposed fix

Apply the same pattern as CogitoRenderer: thinking is enabled only when the model supports it AND user explicitly requests it.

Renderer fix

func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) {
    // thinking is enabled: model must support it AND user must request it
    enableThinking := r.isThinking && (thinkValue != nil && thinkValue.Bool())
    // Use enableThinking instead of r.isThinking throughout the method
    // ...
}

Parser fix

func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
    p.tools = tools
    p.setInitialState(lastMessage, thinkValue)  // Pass thinkValue
    return tools
}

func (p *Qwen3VLParser) setInitialState(lastMessage *api.Message, thinkValue *api.ThinkValue) {
    prefill := lastMessage != nil && lastMessage.Role == "assistant"

    // Check both hasThinkingSupport AND thinkValue (same pattern as CogitoRenderer)
    thinkingEnabled := p.HasThinkingSupport() && (thinkValue != nil && thinkValue.Bool())

    if !thinkingEnabled {
        p.state = CollectingContent
        return
    }
    // ...
}

Behavior after fix

think parameter Current behavior Expected behavior
true thinking thinking
false thinking (bug) non-thinking
not specified thinking non-thinking (consistent with Cogito)

Workaround

Change Modelfile to use non-thinking variants:

RENDERER qwen3-vl-instruct
PARSER qwen3-vl-instruct

Then recreate the model:

ollama create your-model -f Modelfile

Environment

  • OS: Any
  • Ollama version: Latest (main branch, also affects released versions)

Related issues

  • #10961 - Enable/disable thinking through modelfiles (feature request)
  • #10964 - think:false output is unstable (different root cause - model behavior)
  • #11712 - qwen3:235b /nothink doesn't work
Originally created by @yuu104 on GitHub (Dec 6, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/13353 ## What is the issue? When using `qwen3-vl` model with `RENDERER qwen3-vl-thinking` (which is the default for the official qwen3-vl model), the `think` API parameter is completely ignored. **Expected behavior**: - `think: true` → thinking mode enabled - `think: false` → thinking mode disabled - `think` not specified → thinking mode disabled (consistent with `CogitoRenderer`) **Actual behavior**: The model always operates in thinking mode regardless of the `think` parameter, resulting in slow responses (50+ seconds). ### API request example ```bash curl http://localhost:11434/api/chat -d '{ "model": "qwen3-vl", "messages": [{"role": "user", "content": "Hello"}], "think": false, "stream": false }' ``` Even with `think: false`, the model still enters thinking mode. ## Root cause (code analysis) ### 1. Renderer issue `model/renderers/qwen3vl.go:74`: ```go func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) { ``` The `thinkValue` parameter is discarded with `_`. The renderer only uses `r.isThinking` which is set at construction time based on the renderer name (`qwen3-vl-thinking` vs `qwen3-vl-instruct`). ### 2. Parser issue `model/parsers/qwen3vl.go:57-61`: ```go func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool { p.tools = tools p.setInitialState(lastMessage) // thinkValue is NOT passed here return tools } ``` `thinkValue` is received but not passed to `setInitialState`, so it has no effect. ## Comparison with working implementation `CogitoRenderer` correctly respects the `think` parameter (`model/renderers/cogito.go:14-20`): ```go func (r *CogitoRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) { // ... // thinking is enabled: model must support it AND user must request it (true) enableThinking := r.isThinking && (thinkValue != nil && thinkValue.Bool()) // ... } ``` This pattern should be applied to `Qwen3VLRenderer` and `Qwen3VLParser` as well. ## Proposed fix Apply the same pattern as `CogitoRenderer`: thinking is enabled only when the model supports it AND user explicitly requests it. ### Renderer fix ```go func (r *Qwen3VLRenderer) Render(messages []api.Message, tools []api.Tool, thinkValue *api.ThinkValue) (string, error) { // thinking is enabled: model must support it AND user must request it enableThinking := r.isThinking && (thinkValue != nil && thinkValue.Bool()) // Use enableThinking instead of r.isThinking throughout the method // ... } ``` ### Parser fix ```go func (p *Qwen3VLParser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool { p.tools = tools p.setInitialState(lastMessage, thinkValue) // Pass thinkValue return tools } func (p *Qwen3VLParser) setInitialState(lastMessage *api.Message, thinkValue *api.ThinkValue) { prefill := lastMessage != nil && lastMessage.Role == "assistant" // Check both hasThinkingSupport AND thinkValue (same pattern as CogitoRenderer) thinkingEnabled := p.HasThinkingSupport() && (thinkValue != nil && thinkValue.Bool()) if !thinkingEnabled { p.state = CollectingContent return } // ... } ``` ## Behavior after fix | `think` parameter | Current behavior | Expected behavior | |-------------------|------------------|-------------------| | `true` | thinking | thinking | | `false` | thinking (bug) | non-thinking | | not specified | thinking | non-thinking (consistent with Cogito) | ## Workaround Change Modelfile to use non-thinking variants: ``` RENDERER qwen3-vl-instruct PARSER qwen3-vl-instruct ``` Then recreate the model: ```bash ollama create your-model -f Modelfile ``` ## Environment - **OS**: Any - **Ollama version**: Latest (main branch, also affects released versions) ## Related issues - #10961 - Enable/disable thinking through modelfiles (feature request) - #10964 - `think:false` output is unstable (different root cause - model behavior) - #11712 - qwen3:235b `/nothink` doesn't work
GiteaMirror commented 2026-05-04 23:19:40 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Dec 6, 2025):

qwen3-vl is not a hybrid model. If you don't want thinking. use the instruct version of the model.

<!-- gh-comment-id:3620006552 --> @rick-github commented on GitHub (Dec 6, 2025): qwen3-vl is not a hybrid model. If you don't want thinking. use the [instruct](https://ollama.com/library/qwen3-vl:8b-instruct-q4_K_M) version of the model.
GiteaMirror commented 2026-05-04 23:19:41 -05:00
Author
Owner
Copy Link

@yuu104 commented on GitHub (Dec 6, 2025):

@rick-github
Thank you for the clarification. I understand now that qwen3-vl is not a hybrid model like qwen3, and users should choose between the thinking and instruct variants at the model level rather than using the think API parameter.

<!-- gh-comment-id:3620201805 --> @yuu104 commented on GitHub (Dec 6, 2025): @rick-github Thank you for the clarification. I understand now that qwen3-vl is not a hybrid model like qwen3, and users should choose between the thinking and instruct variants at the model level rather than using the `think` API parameter.
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#70877
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?