What do you guys think about Qubes shipped with a stock LLM template?
I am jumping off the GPU passthrough part here ( I am not designing here), but essentially once started we would be only asked which model to download and to start to deploy it as we find convenient.
I could almost bet local LLM will someday fall under 2nd Amendment. 
Take it from here where you wish, when you wish at all.
Do you mean to be installed by default, or just in the templates repo?
I don’t personally use any AI, but would be fine with it being an option if implemented correctly. I think a better implementation would be a detailed community guide rather than a template, which I’m sure already exists somewhere.
What do you guys think about Qubes shipped with a stock LLM template?
I am against it, way to specific. Just do it yourself: Secure AI Inference with Qubes OS: A GPU Passthrough & Ollama Guide
Maybe you can write a salt script for users.
I could almost bet local LLM will someday fall under 2nd Amendment.
Cyberspace stays free 
It would be interesting to elaborate on this, I think.
default.
I would be against it completely then, and I’d imagine most users would be as well.
It would be interesting to elaborate on this, I think.
I don’t want an LLM to be installed by default because I don’t use AI.
Even for users that do use AI, I really don’t think that they would want an LLM installed by default, as they should have the right to choose which LLM to use. But even if they did, I still don’t think that this is the best way to implement it. As I said in my first reply a community-based tutorial would work best here IMHO, looks like @Schnur provided a good one in their reply.
The best template is one you customize for yourself, not someone else spoon-feeding you one.
Exactly. I would be very happy to customize the template I got by default, like I do with all other default templates I am getting when installing Qubes. But, that’s off-topic?
My statement addresses the proposal directly, stock templates are already decided for you, so I cannot agree with it.
Legit, but self-centered? I already have LLM qube, so I am not speaking on my behalf here, but from the point “what and how it would benefit all users”
As they now have the same right to choose Fedora or Debian and (sys-)Whonix, right?
After successful GPU passthrough, it can be any tutorial out there. A little bit is essentially Qubes-related after that, right?
The issue with your response is assuming ‘stock’ means ‘final’?
We already have stock templates decided for us by the developers, yet we still customize them heavily or not using some of them at all.
On the other hand, there is fallacy with “spoon-fedding” logic, because you assume that every user has the time, expertise, and motivation to build a secure LLM environment from scratch. Meaning, looking at this with “your own eyes”.
I do not conflate different definitions nor impose implicit assumptions, my statement is explicit and self-contained.
I wasn’t trying to be self-centered, but you ask me to elaborate on why I didn’t like the idea. I even added why somebody with a different perspective from mine would still most likely agree with me. I don’t think it would benefit all users, or any users for that matter.
Yes, all though I don’t think it is fair to compare operating systems to an LLM, they’re completely different.
I’m sorry, I don’t understand what you mean here, could you elaborate please?
That’s not a response to my argument — it’s just saying I’m wrong without explaining why. You’re claiming I’m misinterpreting you, but you never actually clarified your definition of ‘stock.’ In Qubes, all stock templates are base that users customize. That’s how system works. A stock AI template would work exactly the same way — starting point, not a forced solution.
And that would be pioneering as I see it. No other OS currently as to my knowledge offers this level of secure, local-by-default AI support out of the box as Qubes could.
Local LLMs are inevitable. Users will go to places where they don’t have to be experts in order to set it up and start to use it. And offering such a template would not be the same as an offering of a “Windows” template as a trick to attract new users
tl;dr: no, thanks.
And then…
I am inviting you to read the topic, where I was clear where I am coming from. Otherwise, any idea can be rejected by the logic used in your response.
Also, you didn’t elaborate why you are against it.
I do not need a starting point offered to me, so there is no value proposition to weigh.
All the answers so far are from individual own perspectives. I admit I expected community awareness and belonging to: would it be good for others (except @blitz who showed empathy to a degree, but maybe didn’t asses other cases where it would be good for others who might want to use it).
It’s good. We learn of each others this way too.
We can only speak for ourselves, and your proposal comes from the same position. The only difference is the public feedback before a GitHub issue is created, a development cycle using Qubes Builder v2, and a pull request upstream. Community awareness works both ways, and the reception so far is irrefutable.