LM studio on Linux fails to load model Qwen3.5-9B

[yoavlevy]

Which version of LM Studio?
LM Studio 0.4.6

Which operating system?
Linux

What is the bug?
LM studio fails to load Qwen3.5 . I tried to load several models. few 9B and few below.

Screenshots

Logs
Add any relevant logs.

To Reproduce
try to load Qwen3.5-9B

[IsaiahSama]

I can confirm, I am getting a similar error, with the output:

🥲 Failed to load the model

Failed to load model

error loading model: error loading model architecture: unknown model architecture: 'qwen35'

Which version of LM Studio?
LM Studio 0.4.6 (AppImage)

Which operating system?
Linux: Pop OS 24.04

What is the bug?
LM studio fails to load Qwen3.5 with error: "error loading model: error loading model architecture: unknown model architecture 'qwen35'"

To Reproduce
try to load Qwen3.5-9B

[SINAPSA-IC]

Likewise in v0.3.39 on Windows 11.

By all accounts (unknown architecture) it is something independent of the OS.

Truly disappointed by the lack of preoccupation of solving the TENS of small bugs left behind since 0.4.0, propagated up to 0.4.6. Because of that, I use this program in its 0.3.39 version. The GUI and ongoing inference/downloads can be abruptly Closed with no confirmation message, by simply clicking on the X on the titlebar (seen this 25 years ago, in Windows 3.1 programs, or was it 35 years) the space in the installation path creates problems (You Only use spaces in paths when You know how to manage them,) cannot Delete a conversation properly or at all, cannot Update via the GUI (then why the option to do so,) the command icons (no text version of them anymore) for Models etc are scatterred all over the place on the UI with no rhyme or reason or method - using v0.4.0+ became a frustrating experience, like a 1-man army of junior intern programmers were handed the Development of an otherwise important piece of software which they destroyed instead, just in spite, or b/c of lack of elementary protocols. /sorry for the rant, but it's scandalous what this program has become, despite the rather Low expectations of the user base. IMO

[IsaiahSama]

Update:

Turns out the issue with architecture, is because of the version of llama.cpp LM studio is using.
To resolve:

1. Open LM Studio
2. Go to settings
3. Under the System tab, open Runtime
4. Search for updates / Update All

@SINAPSA-IC

With regards to the "Fail to load model", I get the same issue if a model is already running. So try going to the "Chat" tab, and ejecting any model you have loaded in.

@yoavlevy

[zessu]

Try restarting your computer after an update

[Nerlight]

The 'qwen35' architecture name is still highlighted in white in model list, indicating it's unrecognized. Same goes with new 'ministral3' series, even though LM Studio and Vulkan runtime are up to date (0.4.6 and 2.5.1 accordingly).

Seems like it's not actually supported yet, despite somebody saying otherwise.

[will-lms]

Thank you for reporting. Both architectures should be supported by the 2.5.1 llama.cpp engine version. I am able to verify support when running on my machine here.

If you are able to help us debug the issue when you attempt to load a Qwen 3.5 or Ministral 3 model (or preferably, both), I'd appreciate you taking the following steps. These logs will help us understand the issue you're hitting.

1. In a terminal window, run lms log stream -s runtime > lms-logs.txt
2. Then, go into the app. Load the model with default settings
3. Run a prompt through the app.
4. Kill the lms log stream ... command and send the lms-logs.txt file.

Thank you for your help!

[yaanlpc]

try the beta cuda llma 2.5 in runtime.

[Momix-77]

I tried heretical-qwen3.5-9b and had no problems with them on LM Studio 0.4.7 beta 2 and vulkan runtime 2.6.0.

[davizp]

I updated the LM Studio to the latest version and fixed the issue for me.

[D1m7asis]

Updating fixed for me too

[patrickaclark]

I didn't have to update or revert CUDA models, but I did have to ensure under the 'Model Defaults' setting, that I changed my Default Context length off the 'Model Maximum' setting. I have 32GB RAM and a 10GB VRAM RTX 3080, and was able to set a custom value of 24576 and it loaded fine but crashed at 32768. So it seems there is something wrong with how it's applying the 'advertised' max context, or it is truly trying to force the max(262144), which is probably way to large for most people's hardware.

[wahmedswl]

this is also happening with gemma4 12b-qat

[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 LlamaV4::load called with model path: /home/waqas/.lmstudio/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/gemma-4-12B-it-QAT-Q4_0.gguf
LlamaV4::load config: n_parallel=4 n_ctx=4096 kv_unified=true
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 gemma4 architecture detected, using AUTO flash attention mode
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.00.109.289 I srv    load_model: loading model '/home/waqas/.lmstudio/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/gemma-4-12B-it-QAT-Q4_0.gguf'
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.00.856.951 W load: control-looking token:    212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.00.857.944 W load: control-looking token:     50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.00.912.437 W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.01.907.510 W llama_context: n_ctx_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.01.932.036 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.02.407.797 E clip_init: failed to load model '/home/waqas/.lmstudio/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/mmproj-gemma-4-12B-it-QAT-BF16.gguf': load_hparams: unknown projector type: gemma4uv
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 0.02.407.845 E mtmd_init_from_file: error: Failed to load CLIP model from /home/waqas/.lmstudio/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/mmproj-gemma-4-12B-it-QAT-BF16.gguf

0.02.407.850 E srv    load_model: failed to load multimodal model, '/home/waqas/.lmstudio/models/lmstudio-community/gemma-4-12B-it-QAT-GGUF/mmproj-gemma-4-12B-it-QAT-BF16.gguf'
[DEBUG] llama.cpp-linux-x86_64-nvidia-cuda12-avx2@2.17.0 (llama.cpp) google/gemma-4-12b-qat pid=14492 [LLMProcess] Failed to load model _0x3ed9d0 [Error]: Failed to load model.
    at _0x5532d2.loadModel (/home/waqas/.lmstudio/llmster/0.0.15-2/.bundle/lib/llmworker.js:1:518350)
    at process.processTicksAndRejections (node:internal/process/task_queues:104:5)
    at async _0x5532d2.handleMessage (/home/waqas/.lmstudio/llmster/0.0.15-2/.bundle/lib/llmworker.js:1:510421) {
  cause: 'Failed to load model',
  suggestion: undefined,
  errorData: undefined,
  data: undefined,
  displayData: undefined,
  title: 'Failed to load model.'
}