github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #14418] qwen3.5:35b issue #55872

New Issue
Closed
opened 2026-04-29 09:50:09 -05:00 by GiteaMirror · 5 comments
GiteaMirror commented 2026-04-29 09:50:09 -05:00
Owner
Copy Link

Originally created by @gemlincong-dotcom on GitHub (Feb 25, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14418

What is the issue?

I have encountered the following issue when using qwen3.5:35b to replace the qwen3:30b model for RAG:
environment: ollama -v
ollama version is 0.17.1-rc1

  1. cannot disable thinking although I send api request with think=False .
  2. although I suggest the model to answer me in Chinese and wrote in the prompt, it still answer me in English.
  3. GGML_ASSERT(ctx->mem_buffer != NULL) failed

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @gemlincong-dotcom on GitHub (Feb 25, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14418 ### What is the issue? I have encountered the following issue when using qwen3.5:35b to replace the qwen3:30b model for RAG: environment: ollama -v ollama version is 0.17.1-rc1 1. cannot disable thinking although I send api request with think=False . 2. although I suggest the model to answer me in Chinese and wrote in the prompt, it still answer me in English. 3. GGML_ASSERT(ctx->mem_buffer != NULL) failed ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-29 09:50:09 -05:00
GiteaMirror commented 2026-04-29 09:50:15 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Feb 25, 2026):

Server logs will aid in debugging.

<!-- gh-comment-id:3958602933 --> @rick-github commented on GitHub (Feb 25, 2026): [Server logs](https://docs.ollama.com/troubleshooting) will aid in debugging.
GiteaMirror commented 2026-04-29 09:50:16 -05:00
Author
Owner
Copy Link

@gemlincong-dotcom commented on GitHub (Feb 25, 2026):

Server logs will aid in debugging.

regardless issue 2 and 3, about issue 1, seems the qwen3.5 disable thinking is different with qwen3

Image
<!-- gh-comment-id:3960161341 --> @gemlincong-dotcom commented on GitHub (Feb 25, 2026): > [Server logs](https://docs.ollama.com/troubleshooting) will aid in debugging. regardless issue 2 and 3, about issue 1, seems the qwen3.5 disable thinking is different with qwen3 <img width="603" height="883" alt="Image" src="https://github.com/user-attachments/assets/6a3090da-69da-4f0d-bff5-76182742f439" />
GiteaMirror commented 2026-04-29 09:50:16 -05:00
Author
Owner
Copy Link

@rocaltair commented on GitHub (Feb 26, 2026):

Have you noticed that with same activation parameters under thinking mode, Qwen3.5:35B-A3B runs at
only half the speed of Qwen3:30B-A3B? @gemlincong-dotcom
Is it due to the model architecture or due to Ollama?

Qwen3:30B-A3B(quantization:Q4_K_M): about 61 tokens/s
Qwen3.5:35B-A3B(quantization:Q4_K_M): about 27 tokens/s

OS

Mac

Hardware

Apple silicon M4Pro with 48GB unified RAM

Ollama version

0.17.1-rc2

<!-- gh-comment-id:3965375198 --> @rocaltair commented on GitHub (Feb 26, 2026): Have you noticed that with same activation parameters under thinking mode, Qwen3.5:35B-A3B runs at only half the speed of Qwen3:30B-A3B? @gemlincong-dotcom Is it due to the model architecture or due to Ollama? Qwen3:30B-A3B(quantization:Q4_K_M): about 61 tokens/s Qwen3.5:35B-A3B(quantization:Q4_K_M): about 27 tokens/s # OS Mac # Hardware Apple silicon M4Pro with 48GB unified RAM # Ollama version 0.17.1-rc2
GiteaMirror commented 2026-04-29 09:50:17 -05:00
Author
Owner
Copy Link

@gemlincong-dotcom commented on GitHub (Feb 26, 2026):

Have you noticed that with same activation parameters under thinking mode, Qwen3.5:35B-A3B runs at only half the speed of Qwen3:30B-A3B? @gemlincong-dotcom Is it due to the model architecture or due to Ollama?

Qwen3:30B-A3B(quantization:Q4_K_M): about 61 tokens/s Qwen3.5:35B-A3B(quantization:Q4_K_M): about 27 tokens/s

OS

Mac

Hardware

Apple silicon M4Pro with 48GB unified RAM

Ollama version

0.17.1-rc2

have not compared the generation speed, but qwen3.5:35b is 4 GB larger than qwen3:30b, maybe that affect the speed.

<!-- gh-comment-id:3965830613 --> @gemlincong-dotcom commented on GitHub (Feb 26, 2026): > Have you noticed that with same activation parameters under thinking mode, Qwen3.5:35B-A3B runs at only half the speed of Qwen3:30B-A3B? [@gemlincong-dotcom](https://github.com/gemlincong-dotcom) Is it due to the model architecture or due to Ollama? > > Qwen3:30B-A3B(quantization:Q4_K_M): about 61 tokens/s Qwen3.5:35B-A3B(quantization:Q4_K_M): about 27 tokens/s > > # OS > Mac > > # Hardware > Apple silicon M4Pro with 48GB unified RAM > > # Ollama version > 0.17.1-rc2 have not compared the generation speed, but qwen3.5:35b is 4 GB larger than qwen3:30b, maybe that affect the speed.
GiteaMirror commented 2026-04-29 09:50:18 -05:00
Author
Owner
Copy Link

@ddarmon commented on GitHub (Mar 1, 2026):

I had the same issue 1 as you (i.e., think=false did not always work):

The root cause appears to be that the qwen3.5:35b-a3b model in the registry as of last week had its config set to RENDERER qwen3-vl-thinking / PARSER qwen3-vl-thinking instead of RENDERER qwen3.5 / PARSER qwen3.5.

The qwen3-vl-thinking renderer does not have emitEmptyThinkOnNoThink enabled, so when think=false is set, the assistant prefill is just a bare <|im_start|>assistant\n with no empty <think>\n\n</think>\n\n block to suppress thinking. The model, since it was trained to think by default, starts thinking on its own.

The qwen3.5 renderer/parser handles this correctly — it emits the empty think block as a prefill to tell the model to skip thinking.

Fix: The registry has already been updated with the correct config ("renderer": "qwen3.5", "parser": "qwen3.5"). Re-pulling the model picks up the fix:

ollama pull qwen3.5:35b-a3b

I have confirmed that this works on v0.17.4 after re-pull.

<!-- gh-comment-id:3980537841 --> @ddarmon commented on GitHub (Mar 1, 2026): I had the same issue 1 as you (i.e., `think=false` did not always work): The root cause appears to be that the `qwen3.5:35b-a3b` model in the registry as of last week had its config set to `RENDERER qwen3-vl-thinking` / `PARSER qwen3-vl-thinking` instead of `RENDERER qwen3.5` / `PARSER qwen3.5`. The `qwen3-vl-thinking` renderer does **not** have `emitEmptyThinkOnNoThink` enabled, so when `think=false` is set, the assistant prefill is just a bare `<|im_start|>assistant\n` with no empty `<think>\n\n</think>\n\n` block to suppress thinking. The model, since it was trained to think by default, starts thinking on its own. The `qwen3.5` renderer/parser handles this correctly — it emits the empty think block as a prefill to tell the model to skip thinking. **Fix:** The registry has already been updated with the correct config (`"renderer": "qwen3.5"`, `"parser": "qwen3.5"`). Re-pulling the model picks up the fix: ``` ollama pull qwen3.5:35b-a3b ``` I have confirmed that this works on v0.17.4 after re-pull.
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 bug
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#55872
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?