When you guys use gguf files in ollama, do you normally create a modelfile to go with it, or just hope that whatever default ollama has work with the new model?
https://github.com/ollama/ollama/blob/main/docs%2Fmodelfile....
Xiaomi MiMo Reasoning Model
91–100 of 203 posts
Re: Xiaomi MiMo Reasoning Model
#92Earlier quoted context omitted.
The smaller models have been creeping upward. They don't make headlines because they aren't leapfrogging the mainline models from the big companies, but they are all very capable. I loaded up a random 12B model on ollama the other day and couldn't believe how good it competent it seemed and how fast it was given the machine I was on. A year or so ago, that would have not been the case.
exactly, it seems to validate my assumption from some time ago, that we will mostly use local models for everyday tasks.
strong agree
my employer talks about spending 10s of millions on AI
but, even at this early stage, my experiments indicate that the smaller, locally-run models are just fine for a lot of tech and business tasks
this approach has definite privacy advantages and likely has cost advantages, vs pay-per-use LLM over API.
Re: Xiaomi MiMo Reasoning Model
#93Earlier quoted context omitted.
Xiaomi can also mean millet. I think it's a reference to this Mao quote: https://en.wikipedia.org/wiki/Millet_plus_rifles?wprov=sfla1
That's just a fun coicidence but in reality LeiJun and 12 others from Kingsoft Corp founded Xiaomi after they had a bowl of millet gruel. https://www.scmp.com/abacus/tech/article/3028654/documentary...
This is one of the things that everyone gets the reference, but it won't be good to admit it publicly. This quote is known to almost everyone born in that area, and it's the first thing that come to mind when you hear the name.
Re: Xiaomi MiMo Reasoning Model
#94Its funny to see benchmarks where they omit the top performing models like O3 (Which is the best model in many benchmarks currently) and Gemini Pro/Claude 3.7.
Re: Xiaomi MiMo Reasoning Model
#95Why are there so many English-first AI models from China? Are they not interested in serving their own population? Or is it that if they publish Chinese-first models it won't get publicity in the West?
English won. The Chinese youth struggle to write their own calligraphy characters they can read now. Typing favors English.
The pendulum already turned back. The current generation under 20 grew up with touchscreens. That obseletes input with pinyin; many don't care if the device has no keyboard.
Re: Xiaomi MiMo Reasoning Model
#96Its funny to see benchmarks where they omit the top performing models like O3 (Which is the best model in many benchmarks currently) and Gemini Pro/Claude 3.7.
Those are much much larger models, and they are proprietary. Those model providers just don't have the distilled versions identified and available.
Notice most of the models they are comparing with are 7B models. The exception is also an open weights model (Qwen-2.5-32B-RL-Zero). Even with 32B parameters the MiMo-7B outperforms it.
Re: Xiaomi MiMo Reasoning Model
#97When you guys use gguf files in ollama, do you normally create a modelfile to go with it, or just hope that whatever default ollama has work with the new model? https://github.com/ollama/ollama/blob/main/docs%2Fmodelfile....
If you ollama pull the modelfile will be downloaded along with the blob. To modify the model permanently, you can copypasta the modelfile into a text editor and then create a new model from the old modelfile with the changes you require/made.
Here is my workflow when using Open WebUI:
1. ollama show qwen3:30b-a3b-q8_0 --modelfile
2. Paste the contents of the modelfile into -> admin -> models -> OpenwebUI and rename qwen3:30b-a3b-q8_0-monkversion-1
3. Change parameters like num_gpu 90 to change layers... etc.
4. Keep | Delete old file
Pay attention to the modelfile, it will show you something like this: # To build a new Modelfile based on this, replace FROM with: # FROM qwen3:30b-a3b-q8_0 and you need to make sure the paths are correct. I store my models on a large nvme drive that isn't default ollama as an example of why that matters.
EDIT TO ADD: The 'modelfile' workflow is a pain in the booty. It's a dogwater pattern and I hate it. Some of these models are 30 to 60GB and copying the entire thing to change one parameter is just dumb.
However, ollama does a lot of things right and it makes it easy to get up and running. VLLM, SGLang, Mistral.rs and even llama.cpp require a lot more work to setup.
Re: Xiaomi MiMo Reasoning Model
#98When you guys use gguf files in ollama, do you normally create a modelfile to go with it, or just hope that whatever default ollama has work with the new model? https://github.com/ollama/ollama/blob/main/docs%2Fmodelfile....
I’ll typically use the defaults initially and then use a Modelfile if it’s something I plan on using. I think you can dump the modelfile ollama uses to have a template to work with.
Re: Xiaomi MiMo Reasoning Model
#99Earlier quoted context omitted.
But who will keep them updated and what incentive they would have? That's I can't imagine. Bit vague.
Eventually? Microsoft and Copilot, and Apple and Siri - even if they have to outsource their model making. It will be a challenge to desktop Linux.
I figure this will take the same shape as package distribution. If you have ever used a linux distribution you’ll always see a couple .edu domains serving you packages. Big tech might be able to have specialized models, but following the linux paradigm, it will likely have more cutting edge but temperamental models from university research
Re: Xiaomi MiMo Reasoning Model
#100Umm wow. Great benchmarks. I’m looking forward to chatting with this one.
A couple things stand out to me — first is that the 7B model is trained on 25T tokens(!). This is Meta-scale training; Llama 4 Maverick was trained on 22T or so. (Scout, the smaller model: 40T).
Second, this is an interesting path to take - not a distilled model or an RL layer to get reasoning out of another model, but a from-scratch RL model with reasoning baked in; the claims seem to indicate you get a lot of extra efficiency per-parameter doing this.
I don’t have experience with Xiaomi models, so I’m cautious about this one until I play with it, but it looks like a super viable local reasoning model from the stats.