Trouble with Qwen 3.5 with LMstudio..

Posted by My_Unbiased_Opinion@reddit | LocalLLaMA | View on Reddit | 14 comments

Has anyone got this to work properly? I have tried official Qwen quants as well as Unsloth using the recommended sampler settings. The model usually either has garbled output or straight up loops. I am currently on the latest LMstudio beta with llama.cpp updated to 2.4.0.

Reply to Post

14 Comments

[-]

Key_Cranberry2757@reddit

People, who lies that solve a problem, who YOU really are? for now available version allama marked LM Studio is 2.7. And this version can't proceed new qwen arcitecture. [Total\_Activity\_7550](https://www.reddit.com/user/Total_Activity_7550/) why you can update it and i can't because it is very complex task for common user? [QuirkyDream6928](https://www.reddit.com/user/QuirkyDream6928/) im happy for u if it is true [InevitableArea1](https://www.reddit.com/user/InevitableArea1/) yeah and i bite my butt and now all works fine. If someone get solutions, please stay breef "how to" please.

[-]

Total_Activity_7550@reddit

This is llama.cpp backend not updated in LMStudio. I update it a few hours ago with plain llama.cpp, and now it works better. If you are stuck to LMStudio, wait for an update, or update llama.cpp in settings.

[-]

nderstand2grow@reddit

or choose "unstable" llama.cpp backend in lmstudio

[-]

OsakaDon@reddit

Where is this "unstable" option, I can't find it, I only see "beta".

[-]

My_Unbiased_Opinion@reddit (OP)

You can swap the LMstudio llama.cpp with the latest official llama.cpp github version?

[-]

InevitableArea1@reddit

I kept getting an error with the default prompt templates when using rag. Had to change it myself, just removed {%- if ns.multi_step_tool %} {{- raise_exception('No user query found in messages.') }} {%- endif %} From template and it started working.

[-]

One-Pass3382@reddit

lmstudio 0.4.6 fixed this issue.

[-]

OsakaDon@reddit

I'm on Windows 11 and LMstudio 0.4.6 and I get sudden model unloaded still.

[-]

jbak31@reddit

still not working on lm studio 4.6... linux

[-]

MomentTimely8277@reddit

for me it just stops in the middle of the "think" process. lm studio 0.46

[-]

Significant_Fig_7581@reddit

It works for me... there was an update when the model was just released go check for it

[-]

eworker8888@reddit

If you can install the model on Ollama or Docker Desktop, then you can always use it from E-Worker [https://app.eworker.ca](https://app.eworker.ca) If you just want to test it, then no need to download, just link E-Worker to Open Router (if the model is there) and test directly. No install needed (Web App / Desktop)

[-]

QuirkyDream6928@reddit

working fine for me M4-pro 48GB

[-]

Murgatroyd314@reddit

Both 35B A3B (Staff Pick version, GGUF, Q6) and 27B dense (MLX from mlx-community, 6-bit) are working fine in LM Studio on my M3 Mac.