A little personal project of mine: an AI assistant that runs locally

I’d been toying with the idea of ​​creating an AI assistant that ran locally on my computer for a while, since I already use lmstudio with a few models.

But I also wanted it to have a “3D” avatar and be able to talk.

So, by researching and assisting with an AI (ironic, right?), in my case Gemini, I more or less succeeded this evening.

Anyway, I wanted to make it as simple as possible, especially since I’m no expert in AI, graphics, or anything else. So I found the simplest means to achieve my goal, more or less.

Of course, it lacks complex animation, and in my original intent (but more difficult to achieve in reality), I would have liked to have a personal 3D animated AI assistant that moved around a room, connected to a lmstudio model running locally.

The complex part is missing, but the “brain” runs locally via lmstudio. I chose a small model to avoid overloading the graphics card and leave room for other things like recording video.

For the “environment,” I used Silly Tavern and Silly Tavern Extra (for the voice). For the avatar, I got it from VRoid Hub, since there are some ready-made ones, and so far, they’re the only ones that work for me (I tried other avatars from elsewhere, but they didn’t display; I think the .vrm file is incomplete somehow).

Well, let’s just say I’m happy with that for now.

Maybe in the next few days I can describe what I did better, if anyone’s interested, and I’ll add it to the open post.

Anyway, here’s a quick goodbye screenshot (I don’t think you can upload videos from your computer, so you’ll have to settle for the screenshot XD).

Sounds like an interesting project! I’ll probably dive deeper into AI models after I get my Framework desktop!

What are the steps to create the 3D model and linking it to the “brain”?

All you need is good hardware. I have an AMD Ryzen 7 5700X, 64 GB of RAM, and an AMD Radeon RX 6750 XT (12 GB vram). It’s not bad, but it’s not top-of-the-line either.

You can run a lot of things with this setup.

You just need to find the model. Using LM Studio, you can download it directly from the program, where it also tells you which models are being used entirely by your graphics card and which aren’t.

Then, just put LM Studio in “server” mode with the model you want to use as the “brain.”

To avoid overloading my graphics card, I chose a small model, namely llama-3.2-3b-instruct, because it uses a more “natural” language—if you want a more “human” feel, let’s say.

But I could have also chosen the 8-billion-parameter model. It would have worked just as well, but there would have been a risk that it would have used the PC’s resources (RAM and CPU) and would have slowed down. So I chose the smaller one. It works just as well, and you can continue doing other things in the meantime.