Cannot disable reasoning in Gemma 4 on PocketPal 🙏🏾

[perrypink]

Hello sir 🙏🏾
I’m currently using Gemma 4 (GGUF, Bartowski quantization) on PocketPal, and I’m facing an issue where reasoning (thinking mode) is always enabled.
I tried to disable it using a system prompt (--reasoning off, enable_thinking=False), but it does not work in PocketPal. The model always produces thinking outputs.
This behavior also happens in LM Studio when loading the same GGUF files. It seems that reasoning is always forced on and cannot be turned off.
Expected behavior:
There should be a way to disable reasoning (thinking) via system prompt.
It would be great to have a “Thinking” toggle for Gemma 4, similar to Qwen3.5 in PocketPal, so users can turn on/off.
Additional notes:
enable_thinking=False and --reasoning off does not work.
Issue reproducible on both PocketPal and LM Studio
Gemma 4 is very good, but forced reasoning makes it less practical for simple tasks ☹️
Please update your app soon.
Thank you sir 🙏🏾

[BitaEver]

I have exactly the same problem and it's quite annoying.

[a-ghorbani]

which app version are you using?

[perrypink]

which app version are you using?

PocketPal 1.13.3 (Google Play) sir 🙏🏾

[perrypink]

This is official yaml file for Gemma 4 E2B in LM Studio 🙏🏾

model: google/gemma-4-e2b
base:

• key: lmstudio-community/gemma-4-e2b-it-gguf
  sources:
  • type: huggingface
    user: lmstudio-community
    repo: gemma-4-E2B-it-GGUF
    config:
    operation:
    fields:
  • key: llm.prediction.temperature
    value: 1.0
  • key: llm.prediction.topPSampling
    value:
    checked: true
    value: 0.95
  • key: llm.prediction.topKSampling
    value: 64
  • key: llm.prediction.reasoning.parsing
    value:
    enabled: true
    startString: "<|channel>thought"
    endString: "<channel|>"
    customFields:
• key: enableThinking
  displayName: Enable Thinking
  description: Controls whether the model will think before replying
  type: boolean
  defaultValue: true
  effects:
  • type: setJinjaVariable
    variable: enable_thinking
    metadataOverrides:
    domain: llm
    architectures:
  • gemma4
    compatibilityTypes:
  • gguf
    paramsStrings:
  • 5.1B
    minMemoryUsageBytes: 4200000000
    contextLengths:
  • 131072
    vision: true
    reasoning: true
    trainedForToolUse: true

[a-ghorbani]

A fix is underway.
It is working for some Gemma-4 GGUFs: E.g. https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF/blob/main/gemma-4-E2B-it-UD-IQ2_M.gguf

[dreamsmaner]

I'm having the same issue. Tried various methods to stop thinking on Gemma-4 but nothing worked.
PocketPal Version: iOS v1.13.3 (136)

[a-ghorbani]

v.1.14.0 release resolves this.

25716A76-F5D4-400A-858A-BBF4839EF494.MP4

[dreamsmaner]

How long will it take for the Apple App Store to show this update? Curerntly it still shows v1.13.3.

[a-ghorbani]

How long will it take for the Apple App Store to show this update? Curerntly it still shows v1.13.3.

It has just been approved and published. I'm not sure how iOS distributes it across regions, but it should be available within minutes or hours.

[dyyxxhh]

when using unoffical gemma4 the problem also exist

[a-ghorbani]

when using unoffical gemma4 the problem also exist

could you please share the link to the model? Typically this happens when the template doesn't cooperate.

[dyyxxhh]

when using unoffical gemma4 the problem also exist

could you please share the link to the model? Typically this happens when the template doesn't cooperate.

https://huggingface.co/NullpoLab/gemma-4-E4B-it-Heretic-ARA-Refusals3
i think is this(gemma-4-E4B-it-Heretic-ARA-Refusals3-Q4_K_M.)