LMStudio: Failed to load model error loading model: error loading model architecture: unknown model architecture: 'dots1'
Is anyone getting an error in LMStudio:
Failed to load model error loading model: error loading model architecture: unknown model architecture: 'dots1'
? This is trying to load model quants dots.llm1.inst-UD-IQ2_M.
I assume I need an engine and or LMStudio merge of support for it. And I should just wait and re-try from time to time whether support was merged. Posting here jic it's something else.
My versions on MacOS are
LMStuidio: You are all up to date! The current version is 0.3.16
llama.cpp: Metal llama.cpp v1.33.0
MLX: LM Studio MLX v0.17.0
LMStudio is not working for me either. I don't even see an error log. I just click load and nothing happens. I assume LMStudio needs to be updated before we'll be able to run dots. I'm on windows 11.
It works in llama.cpp so yes, you will need to update LM Studio
There’s no update to support this yet in LM studio.
It's working now after updating llama.cpp engine
Thanks! Updated LMStudio now:
You are all up to date! The current version is 0.3.17
But still, the llama.cpp runtime bundled with todays LMStudio on my MacOS is from a month ago:
Metal llama.cpp
v1.33.0
Engine
Apple Metal accelerated llama.cpp engine
Latest Version Installed
- Llama 4 vision support
- Enable any LLM to be used as an embedding model (requires LM Studio 0.3.16-b6)
- Fixed prompt processing bugs when chats exceed context length
- llama.cpp updated to b5459 (commit 8a1d206)
I see the 8a1d206 commit in llama.cpp be from 2025-05-22 :
8a1d206f [email protected] 2025-05-22 22:21:07 +0300 : tts : fix n_ubatch + make WavTokenizer cache-less (#13713)
I see latter commit from 2025-06-15 that looks like important to have:
9ae4143b [email protected] 2025-06-15 00:52:06 -0700 : model : add dots.llm1 architecture support (#14044) (#14118)
I guess will just wait some more, from LMStudio to move to news llama.cpp. Atm on trying to load dots.llm1 I'm still getting:
error loading model: error loading model architecture: unknown model architecture: 'dots1'
and that kind of makes sense. I can confirm that llama.cpp built from source today works (mbp m2 96gb ram) - this:
sudo sysctl iogpu.wired_limit_mb=80000
build/bin/llama-server --model models/dots.llm1.inst-UD-TQ1_0.gguf --temp 0 --top_p 0.95 --min_p 0 --ctx-size 4096 &
(Edit) Can use larger context with flash attention on - this takes 75gb (v)ram and delivers 16 tps on M2:
build/bin/llama-server --model models/dots.llm1.inst-UD-TQ1_0.gguf --temp 0 --top_p 0.95 --min_p 0 --ctx-size 32758 --flash-attn --cache-type-k q8_0 --cache-type-v q8_0 &
LMStudio updated just now (for Gemma), and got the update for dots.llm1 in too.
Metal llama.cpp
v1.37.1
Apple Metal accelerated llama.cpp engine
- Google Gemma3n support (text-only)
- dots.llm1 model support
- NeoBERT embedding model support
- llama.cpp updated to b5757 (commit bf5bcd0)
Nice! - all should works now. :-)