github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 16:40:08 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14798] qwen3-vl:8b missing thinking toggle template (think: false ignored) #35318

New Issue
Closed
opened 2026-04-22 19:45:33 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-22 19:45:33 -05:00
Owner
Copy Link

Originally created by @tankbottoms on GitHub (Mar 12, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14798

Description

The qwen3-vl:8b model ships with a bare {{ .Prompt }} template that lacks the $.IsThinkSet / $.Think thinking-control logic present in the qwen3:8b model template. As a result, "think": false in API calls is silently ignored for qwen3-vl:8b, while it works correctly for qwen3:8b.

qwen3-vl:8b template (current)

{{ .Prompt }}

No ChatML structure, no thinking toggle -- just a raw prompt passthrough.

qwen3:8b template (correct)

Full ChatML with $.IsThinkSet, $.Think, /think and /no_think toggle, and proper <think> block handling.

Impact

When using the chat API with "think": false, qwen3-vl:8b ignores the flag entirely. All tokens are consumed by thinking output, producing empty actual responses for any non-trivial prompt.

Example: A 4096 num_predict token budget produces 17,054 characters of thinking and 0 characters of response content.

Steps to Reproduce

1. Confirm the template difference

# qwen3-vl:8b -- bare template
curl -s http://localhost:11434/api/show -d '{"name":"qwen3-vl:8b"}' | jq -r '.template'
# Output: {{ .Prompt }}

# qwen3:8b -- full ChatML template with thinking control
curl -s http://localhost:11434/api/show -d '{"name":"qwen3:8b"}' | jq -r '.template'
# Output: Full template with $.IsThinkSet, $.Think, /no_think logic

2. Send a chat request with think: false

curl -s http://localhost:11434/api/chat -d '{
  "model": "qwen3-vl:8b",
  "messages": [{"role": "user", "content": "What is 2+2? Answer briefly."}],
  "think": false,
  "stream": false,
  "options": {"num_predict": 4096}
}'

Actual result: The response message.content is dominated by <think>...</think> blocks consuming the full token budget. The actual answer is empty or truncated.

Expected result: With "think": false, the model should skip thinking and return a direct response, as qwen3:8b does.

3. Compare with qwen3:8b (works correctly)

curl -s http://localhost:11434/api/chat -d '{
  "model": "qwen3:8b",
  "messages": [{"role": "user", "content": "What is 2+2? Answer briefly."}],
  "think": false,
  "stream": false,
  "options": {"num_predict": 4096}
}'

This correctly returns a direct answer without thinking blocks.

Workaround

Using raw: true with explicit ChatML formatting, appending /no_think to user messages, and prefilling <think>\n\n</think>\n\n in the assistant turn works correctly:

curl -s http://localhost:11434/api/chat -d '{
  "model": "qwen3-vl:8b",
  "messages": [
    {"role": "user", "content": "What is 2+2? Answer briefly. /no_think"},
    {"role": "assistant", "content": "<think>\n\n</think>\n\n"}
  ],
  "raw": true,
  "stream": false,
  "options": {"num_predict": 4096}
}'

Expected Behavior

qwen3-vl:8b should ship with the same ChatML template structure as qwen3:8b, including full $.IsThinkSet / $.Think thinking toggle support. Both models are Qwen3-family and support the same thinking/no-thinking modes.

Version Info

  • Ollama 0.16.2 (tested on spark-1)
  • Ollama 0.16.3 (tested on spark-2)
  • Models: qwen3-vl:8b, qwen3:8b
  • OS: Linux (ARM64, DGX Spark)
Originally created by @tankbottoms on GitHub (Mar 12, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14798 ## Description The `qwen3-vl:8b` model ships with a bare `{{ .Prompt }}` template that lacks the `$.IsThinkSet` / `$.Think` thinking-control logic present in the `qwen3:8b` model template. As a result, `"think": false` in API calls is silently ignored for `qwen3-vl:8b`, while it works correctly for `qwen3:8b`. ### qwen3-vl:8b template (current) ``` {{ .Prompt }} ``` No ChatML structure, no thinking toggle -- just a raw prompt passthrough. ### qwen3:8b template (correct) Full ChatML with `$.IsThinkSet`, `$.Think`, `/think` and `/no_think` toggle, and proper `<think>` block handling. ## Impact When using the chat API with `"think": false`, `qwen3-vl:8b` ignores the flag entirely. All tokens are consumed by thinking output, producing empty actual responses for any non-trivial prompt. **Example:** A 4096 `num_predict` token budget produces **17,054 characters of thinking** and **0 characters of response content**. ## Steps to Reproduce ### 1. Confirm the template difference ```bash # qwen3-vl:8b -- bare template curl -s http://localhost:11434/api/show -d '{"name":"qwen3-vl:8b"}' | jq -r '.template' # Output: {{ .Prompt }} # qwen3:8b -- full ChatML template with thinking control curl -s http://localhost:11434/api/show -d '{"name":"qwen3:8b"}' | jq -r '.template' # Output: Full template with $.IsThinkSet, $.Think, /no_think logic ``` ### 2. Send a chat request with think: false ```bash curl -s http://localhost:11434/api/chat -d '{ "model": "qwen3-vl:8b", "messages": [{"role": "user", "content": "What is 2+2? Answer briefly."}], "think": false, "stream": false, "options": {"num_predict": 4096} }' ``` **Actual result:** The response `message.content` is dominated by `<think>...</think>` blocks consuming the full token budget. The actual answer is empty or truncated. **Expected result:** With `"think": false`, the model should skip thinking and return a direct response, as `qwen3:8b` does. ### 3. Compare with qwen3:8b (works correctly) ```bash curl -s http://localhost:11434/api/chat -d '{ "model": "qwen3:8b", "messages": [{"role": "user", "content": "What is 2+2? Answer briefly."}], "think": false, "stream": false, "options": {"num_predict": 4096} }' ``` This correctly returns a direct answer without thinking blocks. ## Workaround Using `raw: true` with explicit ChatML formatting, appending `/no_think` to user messages, and prefilling `<think>\n\n</think>\n\n` in the assistant turn works correctly: ```bash curl -s http://localhost:11434/api/chat -d '{ "model": "qwen3-vl:8b", "messages": [ {"role": "user", "content": "What is 2+2? Answer briefly. /no_think"}, {"role": "assistant", "content": "<think>\n\n</think>\n\n"} ], "raw": true, "stream": false, "options": {"num_predict": 4096} }' ``` ## Expected Behavior `qwen3-vl:8b` should ship with the same ChatML template structure as `qwen3:8b`, including full `$.IsThinkSet` / `$.Think` thinking toggle support. Both models are Qwen3-family and support the same thinking/no-thinking modes. ## Version Info - **Ollama 0.16.2** (tested on spark-1) - **Ollama 0.16.3** (tested on spark-2) - **Models:** `qwen3-vl:8b`, `qwen3:8b` - **OS:** Linux (ARM64, DGX Spark)
GiteaMirror added the thinking label 2026-04-22 19:45:33 -05:00
GiteaMirror commented 2026-04-22 19:45:35 -05:00
Author
Owner
Copy Link

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

qwen3-vl comes in two variants, thinking and instruct (non-thinking). If you want to disable thinking, use the instruct model qwen3-vl:8b-instruct.

<!-- gh-comment-id:4046984259 --> @rick-github commented on GitHub (Mar 12, 2026): qwen3-vl comes in [two variants](https://huggingface.co/collections/Qwen/qwen3-vl), thinking and instruct (non-thinking). If you want to disable thinking, use the instruct model [qwen3-vl:8b-instruct](https://ollama.com/library/qwen3-vl:8b-instruct).
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#35318
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?