24GB GPU

Qwen 3.8 ( 2 variants )

Pangolin>OpenWebUI>lama.cpp for general use and available outside the homelab

Zoo Code>OpenWebUI>lama.cpp for coding , mostly scripts and html

Qwen 3.8 was a game changer for me.

I recently gave Strata a shot for my local AI. I’m very impressed with the performance they have eeked out of my hardware. I’ve got a rather beefy gaming desktop I built as an AI server back when the current craze started, before prices started going up, but models up until this point really failed to take full advantage of that hardware. Strata (which uses a Qwen model) both simplified the setup process, and it maxes out the hardware when running.

Strata acts like an OpenAI compatible API server, I then have Hermes as a “harness” to make an “agent”, which I interact with on “buzz”, which is like discord, but made for bots. Although Hermes also supports discord and several other chat platforms.

If you have the hardware to run Strata, I recommend setting it up, it comes with a basic chat interface, which you can then grill to assist you with setting up Hermes and whatever else.

One thing I didn’t mention, is security. I highly recommend you treat AI as unpredictable, unreliable, and capable of being taken over by bad actors. This isn’t a problem for a pure chat interface, but Hermes gives the AI tools to do things, and those tools could be used to take over your computer, scan your network, and create all kinds of havoc. For my part, Hermes is running in a virtual machine which is on an isolated network. I keep tight control on what it does and doesn’t have access to in my network main network. That may not be something you have the hardware or knowledge to do, at the very least I recommend running Hermes in a docker container, and not giving it access to your personal desktop nor credentials to things like your email account.

If you haven’t seen it already, take a look at !localllama@sh.itjust.works

What a timely comment.

I recently gave my kid a gaming PC for his birthday…which I used as an excuse to upgrade my GPU from a 1660 Super to a 5060ti 16GB

Combined with a 5600x3d and 32gb of system memory.

I just started messing with local LLMs yesterday. “Getting started” is rather easy. pacman -S lm-studio (I use arch, btw), launch it, download a model that your hardware supports, and start talking.

Right now I’m working on getting it to talk to vscode.

As a service? No, not really.

I use llms for specific tasks, usually small models, for things like identification and tagging (immich, paperless), translation or transcription (whisper), etc. I have a bunch of larger models on a Mac mini and my work laptop for various testing purposes, and have used them to make summaries and other such things.

They aren’t great for search IMO, and search summaries can be quite bad.

I do find them useful to interrogate an unknown rs232 protocol or similar, though, which is where larger models come in handy.

Corporate LLMs are generally garbage to me not only because they are just data farming, but also because they really aren’t good at everything. They might claim it, but the reality is a tailored model is often the better approach.

Using openwebui as a front end for llama.cpp is a decent strategy, combined with sub-agent routing (so generic model hands off to tailored models then returns the right info). This can be useful for some scenarios.

A caveat on coding - IMO, they are mostly bad at it. If you are using it to generate a small part of something that needs to be repeated they are handy, but at scale they aren’t great, meaning for the entire project (or even an entire plugin or module, depending on complexity). Maintainability becomes a serious issue. While they have improved somewhat, there are limits to what they can do.

Macs with M series chips are good options for a hardware base since the memory is shared, and IMO is one of the most cost effective options over a dgpu with a ton of memory. A dgpu with a decent bit of memory (say 8-12GB) is a solid choice for a model thats larger (27b with 3 or 4b active as a mixture-of-experts model) especially if you have an iGPU for regular graphics tasks. Since its all shared on the M series macs, just get something that fits entirely in memory there.

Hardware is a 5070 12GB VRAM, an AMD 5950x and 64G of DDR4 running on Nobara (Fedora spinoff).

I use 3 models for most things: Ternary Bonsai 27b and Gemma 4 E4B Uncensored Q8_0 for everyday stuff, and Qwen Agentworld 35B A3B Q8_0 with CPU offloading for tasks that require to keep the model on track for long periods of time.

Everyday performance is best with Bonsai for me - it comfortably fits in my VRAM and is noticeably better with staying on task than similarily sized models. But if i want it to work overnight on a task without issue, the lower speed of the agentic Qwen is outweighed by its capability to do this.

They are allowed tool calls for websearch and reading webpages and have a sandboxed directory for R/W operations, but anything regarding command line and R/W access in other areas has to be manually authorized by me (i’m not willing to play with fire here).

They are currently not reachable from outside, but i am currently building a smaller machine as a home server, which will then be using a 3070Ti and 32G RAM to provide probably the Bonsai Model for me on the go per VPN.

The smaller models are here for quickly creating documents i do not care to do myself, RAG for technical documentation, creating quick python scripts for automating small tasks and teaching me RegEx and Python(mostly by analyzing the LLM output and playing around with it) and also some RP as NPC in a fantasy setting. The agentic model is currently working on cataloguing my games collection (around 36000 titles) by looking up games, their genres and descriptions and filling in blanks in the database. (Funnily, this task was refused by any model freely available because of the size of the database. Even the local models moan and bitch that it takes so long. No shit, this is why i outsource this!)

Qwen 3.8 flash next on a 4090, using strata instead of llama.cpp, gives me about 45-55 tokens/sec. Mainly used for coding tasks. The model size is perfect. If the qwen 4.0 models can be a bit less wordy in reasoning but the same size and quality they’ll be really strong.

Played around with it on my desktop machine. Someone’s distilled variant of Qwen 3.8-27B-Q4_K_M, on a intel12700k + RTX3090 24GB is what I use.

I haven’t adjusted the reasoning down from very high but I mostly use it for trying tweaks to half-finished coding projects I’ve lost motivation to complete many years ago. I used it to develop a very simple JS-free personal website. I have secure shell which I can access from Termux or PuTTY, but sometimes I have the llama.cpp webui insecurely accessible online for one day at a time just for fun.

My server’s old GPU (8gb) can’t run anything useful LLW wise I don’t think.

Have run some stuff locally on my main computer, but nothing setup as a proper service.

I self-host my own neural network, all within the confines of my own skull.

Yes, were all on c/LocalLLaMa

I’m running qwen3.8-next-flash on a rx7900xtx plus 128gb of ram. It serves my hermes agent. I direct it through matrix. I also forward an endpoint to my VPS to use for anything. Mostly opencode.

I just got qwn3-4b-instruct working on an Intel A750 with 8 gigs of vram.

I use it to power Home Assostant/Voice Assistant on a few M5Stack echos around the house to replace cloud based smart speakers.

I wouldn’t say I ‘run’ it, but I mess around with it

I have a base level Mac Mini I got before the price increase a few years ago, also from when models were smaller and my ambitions were less. I’ve contemplated upgrading it, but I have other things to spend my money on.

M4 processor, 16 gb ram, 256 gb storage, and a 4tb external HDD drive (an old NAS drive).

It’s fine. It’s mostly a docker machine, but I do throw AI tasks at it a few times a week. Mostly stuff that I just let it chew on for 5-10 minutes and come back to.

I run openwebui, and have several models I bounce between. I mostly run Gemma4 (Google) and Granite4.2:8B (IBM).

Editing to add: the self hosted version of all these models, via llama, just in case that was unclear.

Running Qwen 3.8 27B dense on some R9700s, and Qwen 3.8 flash next on dgx sparks. I use it for code gen outside work, including an automated Dev pipeline that rubs on its own. When I did the math comparing to token API pricing of Opus 4.8, it covered its own costs in a few months. With the sparks I get about 40tg/s, 1500pp/s, but can get concurrency up to 5 stably. With the R9700s, I get ~100tg/s, 3000pp/s and can also get up to 5 concurrency.

I access it over wireguard to the network that it is on, and use termux to ssh into a build machine for dev over CLI using https://pi.dev/

I also started building https://fcli.dev/ to make integrations easier; the idea being that LLMs should just implement simple reusable functions instead of repeatedly making complex curl requests to integrate with a service.

I also use LLMs heavily for work; Qwen 3.8 seems almost on par to Opus 4.8 IMO. I also don’t have to worry about token costs, and it often responds faster than Claude would because there usually isn’t a queue on my machines. At this point, I wouldn’t complain if my work provided me 2x B6000 instead of a Claude subscription; I’d still be able to get all my work done.