My Homelab: Tools and Services I've Self-Hosted

Unlimited Grunt Work Without Token Anxiety

Since building the home server, I've experimented with a variety of tools and services. The hardware itself is only half the fun. The single biggest perk of local large language models (LLMs) over cloud models is stability. No corporation will suddenly change the top-k or temperature or the default thinking level. The models won't get silent updates. I use the local LLM for data cleanup and analysis, and I don't use it for coding. Nearly all of the output I get from it is JSON. It's perfectly fit for this kind of task, where I don't need much creativity or thought, but there's lots and lots of work to be done. I've tried doing the same thing with Claude and locked myself out for a week. With a local LLM, there's no token limit.

The Local LLM

I've benchmarked plenty of the top open-weight local models on this machine. I default to Qwen3.6 27B, which has great performance given its relatively tiny size. But I'm not putting this at the top of the list because I'm completely happy with it. Even with a 24GB card, I can't start seriously replacing cloud AI models with my local one. Qwen3.6 27B is great, but I have to run it quantized, and even then, it takes up a good 21GB of my VRAM. That leaves little room for the cache, so I can't really crank up the context. And what good is AI without context? So, for now, my local LLM really is limited to grunt work. Maybe newer models will get more efficient. Or maybe I just need more VRAM.

ComfyUI

With the RTX 3090's 24GB of VRAM, I can use some really good models. All locally! That gives me the same advantage as the LLM: control and stability. I get to control every parameter, and I can lock everything in to get consistent output instead of hoping a cloud service behaves the same way tomorrow. The output isn't limited to images, either. Videos (I haven't tried those yet), 3D models (mind-blowing), background removal, lighting adjustments, and pretty much everything else. It's especially great because I can build automations around it, and it doesn't cost money every time something makes a request. Unfortunately, there's a catch, and it's the same one as before: VRAM. It's not that I want to run larger image models. The problem is that if I want to run a decent image model, I first have to unload the LLM I was using for other tasks. I can't really have my LLM and image model working at the same time because there just isn't enough space. And because I have lots of data work that depends on the LLM, that's made me use ComfyUI less than I'd like.

Telegram Bots

I've hosted Telegram bots before, but they always ended up being more of a learning project than something useful. The biggest problem was that I couldn't really host them. I'd run them on my computer, and then they'd go down whenever my computer was off. Otherwise, I'd need to rent a VPS. Now I have tons of virtual servers sitting in my basement. It's high time I raised a bot army! What kind of bots?