Self-Hosting My Own LLMs

At a time when ChatGPT, Claude, Gemini and plenty of others are being handed to users for free, why go to the time and expense of recreating something that’s admittedly inferior? I’ll go through a few reasons, but the main one comes down to data sovereignty. I wanted full ownership and control over my own data. Full stop.

I know I’m an outlier here. But it’s the same reason I still pay — in 2026, even! — for an email service instead of relying on Gmail. Throughout 2025 I found myself having more and more conversations across ChatGPT, Claude and other services, and I realized those conversations were deeply personal and valuable to me, even when I wasn’t chatting about anything especially private. I wanted to be able to look back and easily find a conversation I’d had months (or eventually years) earlier. I know ChatGPT, Claude and Gemini all offer that kind of history — but I’d grown uncomfortable with the idea that I was building up a library of my own conversations inside someone else’s walled garden. Technically, they own that data, and therefore my conversations. I felt very little agency over something that felt like it should be entirely mine. So it was settled: I would use open-source tools to host my own chat interface — and if at all possible, power it with large language models running on my own machine.

There were other reasons, too. I like tinkering, and I wanted to see how the LLM sausage gets made. Chatting with ChatGPT online felt mysterious and hard for me to fathom — how did this new technology actually work? There’s no better way to learn than to set it up locally and watch all the moving parts.

Lastly, there’s the self-sufficiency angle. We’ve already seen powerful closed-source models kept behind tightening restrictions and steep prices — Anthropic’s Mythos and Fable, OpenAI’s GPT-5.6. And even when the models aren’t formally restricted, plenty of people swear up and down that the big providers quietly throttle them at busy times. Whether that’s actually true is more or less unknowable — but there’s an antidote either way: use open-weights models that aren’t so directly under any one company’s control. A local model, running on my own hardware, is something that neither the government nor big business can so easily restrict or take away.

Before I begin, I should give some context about the sort of hardware I have. It’s 2026, and hardware — memory especially — is suddenly expensive, because we’re in the thick of a generative-AI bubble. People are paying scalper’s prices for the hardware that lets you run models locally. Pretending this is all doable without dropping a modest amount of money would be disingenuous. So what am I running, and what did it cost?

I’m running what most people would call an enthusiast’s setup: some decent hardware, but nowhere close to a high-end build. I started this with a custom AMD desktop I built right before COVID, paired with a video card I picked up last year. In 2026 it’s not remotely cutting edge, and that’s fine.

I built it around AMD’s Ryzen 9 5900X — a 12-core beast at the time — and loaded it up with about 80 GB of DDR4. I really wish I’d bought more memory back when it was cheap a couple of years ago, but so does everyone else. Even so, this is a very solid (if not cutting-edge) machine.

Last summer I walked into Micro Center and walked out with a refurbished NVIDIA RTX 3090 Ti for $850 plus tax. I remember wondering at the time whether it was a dumb purchase. I’m now sleeping very easy with that acquisition — it turned out to be a great buy. The 3090 Ti and its 24 GB of VRAM do the majority of the heavy lifting in my system, and that 24 GB is the single most important number in this whole story: it decides which models fit, and shapes nearly every configuration choice I make later on. Luckily I’d already put a largish 850-watt power supply in the machine, so I didn’t need to upgrade that either.

So how much would this setup cost to assemble today? This got me curious about my original build cost. Back in Chicago I actually lived within walking distance to a Micro Center, a deeply expensive convenience for me. I did my whole build using Micro Center, and I thought it would be fun to pull up my original receipts…​mostly just to cry at the prices.

I originally built the computer with a Ryzen 5 3900X in it, but I upgraded later, so I’m only listing components in the final build. Most of the machine came together in mid 2020. In Sept 2020 I bought an AMD RX 5500 XT that had 8GB of ram for $200 (also refurbished at Micro Center). At the time, crypto mining was placing an enormous demand on cards, especially Nvidia ones. I wasn’t yet ready to pay blockbuster prices for a super in-demand card, which is how I settled on the RX 5500 XT. Now that I reflect back on the prices I paid, I’m actually a little surprised to see that I dropped $265 on two 32GB sticks of RAM. That’s more than I could buy it for today I think, which puts the current memory bubble into perspective a little. Even though we’re in the midst of global memory shortage and memory of all kinds (current, last generation) have spiked since a year or two ago, my anecdotal prices are still cheaper than they were in 2020. Interesting.

We’ll get into the details on each component below, but first let’s zoom out so we don’t miss the forest for the trees. Here’s the whole stack on one page:

Everything from Open WebUI rightward is self-hosted on hardware I own; only the phone’s traffic takes the Tailscale detour to get in. (As we’ll see, Open WebUI and the inference server don’t even have to live on the same machine.)

My AMD desktop runs Linux. Which distribution probably doesn’t matter much — I’m on a Debian variant at the moment, though I might redo the whole setup on Fedora one of these days (like I said, I like to tinker). This guide assumes a Linux environment.

A quick note on the GPU: I actually got things running first on a small 8 GB AMD Radeon card I already had, before the NVIDIA. It’s certainly possible to use a GPU from someone other than NVIDIA, but I struggled a bit getting the AMD card working — non-NVIDIA cards are still something of a second-class citizen in the local-LLM world. What tripped me up more, though, was that card’s meager 8 GB of VRAM, which is ultimately what pushed me to jump to the 3090 Ti.

So what does the stack actually need? Besides the open-weights models themselves, you need something to perform inference — the act of taking incoming text (my chat prompts) and running it through the model to produce a text response. There are several options here, and I landed on llama.cpp. It’s not as hardcore as vLLM, but it’s a little more "enterprise" than something like Ollama or LM Studio.

By design, though, llama.cpp doesn’t hand you a nice web interface for actually chatting with the model. For that you need a chat front-end. There are many (LibreChat, Oobabooga and SillyTavern all came up while I was looking), but I settled on Open WebUI. A big reason was that I could reach my Open WebUI server from my phone — an important detail for me. I wanted the whole setup to be usable from both my desktop and my phone while I’m on the go, and at the time Open WebUI had the most viable options for use on my iPhone. (Incidentally, I ended up using an iOS app called "Conduit" to reach my Open WebUI server from the phone — though the app ecosystem is always shifting.)

With Open WebUI out front and llama.cpp doing the inference, I was nearly in business. My remaining problem was picking which open-weights model to run — there were actually several intriguing candidates, and I wanted to switch between them easily. Trouble is, changing models would normally mean restarting llama.cpp, and that’s not something I can do from my phone.

That’s where llama-swap comes in. It acts like a little local router that sits between Open WebUI and my llama.cpp inference server, and it lets me expose a handful of different models at once. I pick a model from inside the Open WebUI chat window and just start talking; llama-swap works with llama.cpp to spin up inference on the right model automatically. No stopping and starting services from the command line. It just works.

Lastly, there’s the small matter of reaching all this while I’m out and about. To keep things simple, I took the plunge with Tailscale, which lets you build a personal mesh VPN. I run Tailscale on both my home server and my phone, so when I’m away from home I just make sure I’m connected to my Tailscale network (my "tailnet") and the Conduit app on my phone can reach my Open WebUI server as if I were sitting at home.

I started out by following a few online guides for installing llama.cpp on Linux, and I initially decided to run everything from pre-built Docker containers — it seemed like the cleaner solution, easier to isolate and configure. At first I had separate containers for Open WebUI and llama-swap (whose image also bundled the llama.cpp llama-server binary), and I later added a third for Redis.

Over time, though, I drifted away from containers. The main reason was that I needed to stop using the pre-built images and start running my own custom binaries. For a while I baked those custom binaries into a custom Docker image, but eventually that started to feel like too much work.

The Open WebUI container was the first to go — though it didn’t disappear so much as move. I have a second machine in the house, some sort of Dell OptiPlex, that I use as a souped-up NAS running FreeBSD. It’s a rock-solid server, unlike my AMD desktop, which was constantly crashing while I figured out how to safely load these enormous models onto the 3090 Ti. The 3090 Ti had to stay in the desktop — which dictated where llama.cpp would run — but Open WebUI could technically live anywhere. So I moved it into a jail on the FreeBSD box for the sake of stability.

That move quietly solved a problem I’d been papering over. Back during all the crashing and rebooting, every restart of the Open WebUI service logged my clients out and forced them to re-authenticate immediately — which got old fast. Redis was my fix: it let logins persist across service restarts. But once Open WebUI lived on the rock-solid FreeBSD box, it hardly ever restarted. My inference machine still hiccups and goes down now and then, but that no longer takes Open WebUI down with it. So I just…​ forgot about Redis. I never installed it on the FreeBSD box, and it drifted out of the picture entirely.

That left just one container with llama.cpp and llama-swap in it. For llama.cpp, I kept finding myself wanting the absolute latest version. When Google released the Gemma 4 models, I downloaded one right away and tried to use it — and got errors, because my version of llama.cpp didn’t fully recognize the new model. The fix was to run a newer llama.cpp, but as long as I was using someone’s pre-built Docker image I was at the update mercy of whoever built it. I wanted to be able to pull the latest llama.cpp source, build it myself, and run that. You can do that with Docker, of course — but at some point I decided it was simpler to stop using Docker and just run the binaries directly on Linux. I also wanted to run a slightly modified build of llama-swap (more on that later), and once I was building my own llama-swap binary anyway, wrapping it in a container felt like a pointless extra step.

So today none of this stack runs in Docker. It’s a little simpler for me to manage. Was Docker bad, or slow? Not at all — this really just comes down to preference.

There are several options for serving up the raw language-model files. If you want a GUI and a gentler experience, LM Studio has been around for years and is a great option. But I wanted something that ran from the command line as a service or daemon. Ollama fits that bill, but I kept seeing it described as "llama.cpp with training wheels" — and in that case, I figured, why not just use llama.cpp itself? I also briefly checked out vLLM, but my particular use case — where I’m the sole user of my local stack — didn’t really play to its strengths (vLLM shines when you’re serving lots of concurrent users). llama.cpp turned out to be the sweet spot: a robust headless inference server, with "prosumer" touches like being able to offload layers of a too-big model onto the CPU — at the cost of speed — when VRAM gets tight.

Like I said earlier, I started out running whatever version of llama.cpp happened to ship inside my pre-built llama-swap container. But I wanted to track the very latest release: pull it from GitHub, build it, and run that. That’s ultimately what drove me off Docker and onto binaries installed directly on my Linux system.

The build itself is refreshingly simple. After cloning the repo once, keeping it current and reinstalling looks something like this:

The -DGGML_CUDA=ON flag is what builds in NVIDIA CUDA support. The final sudo make install drops the freshly built llama-server into /usr/local/bin, where the rest of my setup expects to find it.

Going native did give me some heartburn the one time I also tried to run llama.cpp on my AMD card. An AMD card needs llama.cpp built against AMD’s ROCm stack rather than NVIDIA’s CUDA, so I suddenly had two different builds to juggle — one CUDA-compatible, one ROCm-compatible. I managed it by renaming binaries, which worked but felt gross; this is exactly the kind of situation where a Docker container would have been the cleaner approach. In the end, though, I just let the AMD card quietly fade out of my setup. It was only an 8 GB card, after all.

If I’d wanted to just pick one model and stick with it, my life would have been a little simpler. Instead, I wanted to start a chat with one model and then, in the very next session, switch to a different local model. That requires restarting llama.cpp (pointed at the new model each time). Rather than do that by hand, there’s a way to have software restart llama.cpp on my behalf: llama-swap.

llama-swap is a proxy that sits in front of llama.cpp. Requests hit llama-swap first, and it looks at which model was requested. If llama.cpp is already running that model, llama-swap just passes the messages along and llama.cpp infers away. But if the request is for a different model, llama-swap does the dirty work: it stops llama.cpp and restarts it with the newly requested model. There’s a 10–15 second wait while llama.cpp reloads, but after that it’s back to inferring away.

llama-swap looks like pretty simple software, so why did I want to run a custom version of it? My problem was that, early on, llama.cpp was often crashing on me — usually because I’d given it incorrect model settings. Unfortunately, when llama.cpp crashed, llama-swap would essentially swallow the error, and I’d never see anything useful in my chat window. So I wrote a small patch to fix exactly that. It took a little fiddling, but I eventually got it working.

At first I was going to post a diff of my patch, but while preparing it for this post I noticed that the patch no longer works for the latest version of llama-swap. I’m in the process of trying to get a revised patch merged into the llama-swap project itself, so fussing over this may soon no longer be needed. Here is the basic idea behind the patch:

llama-swap already captures the upstream process’s output (both stdout and stderr) into a log buffer. My change hooks the point where llama-swap gives up on a model that exited before it ever became ready: instead of returning a generic "the process exited" message, it pulls that captured output and folds it into the error it hands back. So the client sees the model’s actual complaint — a rejected flag, an out-of-memory abort, whatever it was — rather than a shrug. The one wrinkle is that the log buffer accumulates across attempts, so I also clear it at the start of each launch; a repeated failure then shows only the latest attempt’s output instead of stacking duplicate copies.

For llama-swap to know which models it can offer up, it needs a configuration file. That file is essentially a collection of llama-server startup commands — one for each model you want to make available. Want to run a model with a 128k-token context instead of 64k? Configure that here. Want to run a model half on the GPU and half on the system CPU? That goes here too. Basically any setting that controls how llama.cpp gets launched lives in the llama-swap config file.

So what does mine look like? This isn’t the whole thing, but it’ll give you the idea. (And no, I don’t really know whether these settings are truly optimal — I’ve just tinkered with them until each model runs okay on my machine.)

You’ll notice each command uses ${PORT} rather than a hard-coded port number. llama-swap fills that in when it launches the model, so it stays in control of the wiring between itself and llama.cpp.

Everything up to this point has been plumbing. Open WebUI is the part I actually look at — the ChatGPT-style chat window in my browser and on my phone. As I mentioned earlier, it’s the one piece of the stack that doesn’t run on my AMD desktop; it lives over on the FreeBSD NAS instead.

There’s a small irony here. Open WebUI is really designed to be run as a Docker container — that’s the blessed, easy path. And I run it about as far from that path as you can get: natively, inside a FreeBSD jail, on top of an emulated Linux. It was more work, but it fits how I wanted the rest of the house organized, and it’s been rock solid.

The jail itself is a plain FreeBSD 13.5 jail, created with iocage. The one non-obvious wrinkle is that Open WebUI can’t run on FreeBSD directly, so the jail has FreeBSD’s Linux compatibility turned on — which is what all the linprocfs and linsysfs permissions here are for:

Getting Open WebUI running inside the jail was the fussiest part of this whole project. My first instinct was to install it on FreeBSD’s own Python — and that went badly. Every time I ran pip install open-webui it would grind for a while and then die on some missing build dependency, so I’d install that piece — Rust, then cmake, then OpenSSL, then libffi, then a Postgres client library — rerun pip, and hit the next one. After a couple of hours of whack-a-mole I gave up on the native path entirely.

The fix was to stop fighting FreeBSD and lean on its Linux compatibility instead. I installed a Rocky Linux 9 base (linux_base-rl9) and mounted the Linux proc and sys filesystems that Linux programs expect to find. Then, rather than build a Python myself, I downloaded a prebuilt standalone Linux build of CPython 3.11 (from the python-build-standalone project) and unpacked it into /opt/python. Pointed at pip install open-webui — where ready-made Linux wheels actually exist — that Linux Python installed it cleanly. The whole saga is exactly the kind of thing the official Docker image exists to spare you, which is a little ironic given that leaving Docker behind is what put me here in the first place.

One gotcha worth passing along: those Linux proc/sys mounts have to survive a jail restart, or the Linux Python quietly stops working the next time the machine reboots. I made them permanent by adding the mount to the jail’s fstab from the host:

To keep it running, I wrote a small rc.d service so the jail starts Open WebUI on boot. Under the hood it just runs open-webui serve on port 8080, wrapped in FreeBSD’s daemon for a pidfile and a log. One environment variable in there is worth calling out:

That’s a ten-minute client timeout. With a hosted service, responses come back fast; with a big model running on a single 3090 Ti, a long answer can genuinely take minutes, and the default timeout would hang up on the model mid-sentence. Bumping it to 600 seconds fixed that.

The first thing to do once you can reach Open WebUI in a browser is to point it at the models. Open WebUI talks to inference backends through an OpenAI-compatible API, so in its admin settings I set up a Connection to my local inference server at http://inference-server-ip:9292/v1. That’s llama-swap over on the ryzen desktop — from Open WebUI’s point of view it’s just another OpenAI-compatible endpoint, and behind that address llama-swap and llama.cpp do all the swapping and inferring we set up earlier.

Models are both the fun part and the excruciating part. When I was brand new to hosting a local LLM, the sheer array of open-weights models was disorienting. Where do you even start? This is an area that changes every few months, so anything I write here will start going stale almost immediately. Still, a few guiding principles have stuck with me.

The big one: good models are great, but newer models are probably better. Models are advancing fast right now. If DeepSeek figures out some novel feature or setting, it will likely get studied and reproduced in everyone else’s next wave of models. When I first went looking I searched for "model rankings" and kept reading posts that said things like "Microsoft’s Phi-4 punches above its weight." That’s wonderful — but the Phi-4 family came out in late 2024, and I was assessing the landscape in late 2025, so a lot had changed. Great as models like Phi-4 were when they came out, too much has moved since to stay dogmatically attached to one model like that. So as new models come out, keep your old model files handy, but embrace and test the new ones. They’re probably better — and on the off chance they aren’t, well, you still have the old files. That’s the whole strength of running open-weights models: they’re in my possession, and I can upgrade (or not!) at will. Just since I started, Google dropped its Gemma 4 family, which has been great, and the Qwen models keep coming.

Other open-weights models get released too, but sometimes they’re simply too big for me to run — which brings up size. Figure out your VRAM situation, then pick models that fit it. With 24 GB of VRAM I have a good sense of what I can run and what’s simply out of reach: I want model files in the neighborhood of 15–20 GB, which leaves some GPU memory left over for context (the running size of the conversation). That lands me squarely on models in the 25–35 billion parameter range. These are very strong — capable of genuinely impressive results. Honestly, chatting with a good 30B model isn’t that different from using OpenAI’s or Anthropic’s latest flagships; those are almost certainly far larger, but a typical chat session won’t really showcase the gap. Even dropping into the low teens of billions is a good experience, and smaller models fit more completely inside the GPU, so responses come back snappy.

There are bigger models out there, of course. Some intriguing ~70-billion-parameter models exist, but I can’t come close to running them usefully. And it’s generally believed the true state-of-the-art models live well above that — 200B, 300B, 600B, maybe a trillion or more. None of those are remotely runnable on my machine, even with extreme quantization. So I’ve made peace with sticking to models around 30 billion parameters and under.

Some astute readers might remember that my original motivation for all this was data sovereignty. The pull of running local models on my mighty 3090 Ti was strong — but stronger still was the itch to run the absolute latest, most capable models. I knew I’d never own the hardware for that, so I briefly considered a service like RunPod, which rents out GPU capacity on demand. Wire up a model with RunPod and you’re cooking with gas. I might revisit RunPod someday, but at the time the cost was hard to justify — for me those pricey GPUs sit idle most of the time I’ve got them reserved.

The more cost-effective move was to find someone already selling inference as a service — an inference provider, like the big dogs. Enter OpenRouter. This isn’t a commercial for them, and I’ve heard some misgivings, but for me, creating an OpenRouter account and wiring my Open WebUI up to it instantly expanded my options. Now, starting a chat, I had a choice: talk to my local Qwen model, or reach for one of the latest and greatest models I couldn’t possibly host myself.

Ironically, after all this effort to reproduce something OpenAI and Anthropic give away for free, I’ve found I most often use OpenRouter to push my more demanding conversations through "Sonnet Latest" or "Opus Latest." This time I’m actually paying for that Sonnet or Opus usage — but each conversation costs pennies, or fractions of a penny. Did I actually accomplish anything, then? Hopefully! Because I’m a paying customer rather than a free user, Anthropic and OpenAI have pledged not to use these conversations to train future models. OpenRouter also has a "Zero Data Retention" setting, which I’ve enabled. "ZDR" means the provider actually running the inference has agreed not to store my conversation history for later use (i.e., training). That obviously requires a measure of faith on my part, but I’d rather choose those settings and try to hold providers to them than blindly chat with a free service that makes no privacy claims at all.

The OpenRouter setup with Open WebUI is dead simple. After I created my account and loaded it with $20 or $30 of credits, I generated a one-time token that acts as a sort of password. Open WebUI lets you specify multiple inference endpoints, so alongside the URL for my local services I added OpenRouter’s URL and token. Voilà — the next time I start a conversation, I can pick one of my locally hosted models or any of the huge menu of models on OpenRouter.

I should at least briefly mention Tailscale, because I couldn’t reach this whole local-LLM stack without it. Specifically, I need to get at my Open WebUI server from both home and while I’m out and about.

I suppose I could make Open WebUI publicly accessible and lean hard on its username/password login. But that would mean anyone on the internet could sit around all day probing my server and trying to break in. There’s a better way: keep all my servers private and reach back into my home network through the power of a VPN.

Many years ago I ran an OpenVPN service on my network to connect back to my home servers while away. OpenVPN is fine, but it can be fiddly to set up and a little slow to use. In more recent years a new open-source VPN protocol called WireGuard emerged. Tailscale is a commercial offering — free for personal use — that’s built on WireGuard and makes the whole thing even easier to stand up. I can personally vouch that setting up a Tailscale VPN on my home network was about a thousand times easier than configuring OpenVPN back in the day, and traffic is meant to be faster too, since it’s WireGuard underneath.

So today I have Tailscale running on my FreeBSD machine and on my phone. When I’m off my home wifi, I just stay connected to the Tailscale VPN, and the appropriate traffic gets routed back to my home server over the encrypted link — so the whole setup stays reachable wherever I happen to be.

This is just one implementation, heavily shaped by the current state of open-source software (and, since I’m an iPhone user, the iOS apps that go with it). Options will come and go. I could see replacing Open WebUI with a different front-end if one were compelling enough — though I’d want to bring my chat history along, which would probably mean a custom job to bulk-export it from one and bulk-import it into the next. The fact that I can export and import at all is a strength of open-source software, but it’s also enough of a hassle that it forms a bit of a moat around whatever I’m currently using. Any new front-end would have to be pretty compelling to justify migrating everything over. And they keep appearing — even in the short six months since I built this, I’ve seen front-ends I don’t remember hearing about in late 2025.

Was the effort worth it? I’m very happy with my setup — though I recognize that a desktop plus a capable GPU is a real expense, and not the obvious choice for everyone. If you wanted to roll your own the way I did but without the hardware, a lot of this could be done cheaply in the cloud. Pick a cloud VM provider, run a smallish Linux VM with Open WebUI (or another front-end) on it, and use Tailscale to tie it into your tailnet. If you want on-demand "private" inference, a dedicated GPU cloud provider like RunPod probably makes the most sense. Otherwise, just wire Open WebUI up to OpenRouter — you’ll pay per conversation, but the cost is mere pennies. (AI-assisted coding is where the bills start to add up fast; plain chatting with models is very cheap.) I happen to have a couple of decent-sized computers at home running 24/7, so for me it’s simply cheaper and more convenient to host everything there.

It took me a while — maybe a couple of weeks — to get comfortable with the terminology, the pace of developments (which models are best, how much quantization actually hurts quality, and so on), and the administrative savvy to run my own setup. But once I did, it was great fun to have this much control. More than anything, this exercise showed me that LLMs aren’t some mysterious dark art. Running inference on my own machine was a genuinely satisfying milestone — and it’s within reach for just about anyone, as long as you’re willing to pick a small enough model.