LLM VPS Hosting

LLM hosting means running an open-weight AI model on your own server, where it answers prompts sent to it over the network. Most self-hosted setups use Ollama, a free, MIT-licensed tool that loads models such as Llama 3.2, Qwen3, Gemma 3 or gpt-oss. Ollama is built on the llama.cpp engine, and Open WebUI adds a browser chat on top. Both go onto a GreenGeeks VPS through root SSH, with no Docker involved. Everything runs on the CPU, so the model has to fit in RAM. Speed depends on memory bandwidth more than on how many cores the plan has.

You don't need a GPU for one person using models up to about 14B, and GreenGeeks VPS plans run those on CPU cores. You do need one for real-time chat with several people, for 70B models, or for long documents, since reading a long prompt is slow on a CPU. In 2026 testing, an 8B model ran at 63 tokens a second on an RTX 4000 Ada and about 206 on an H100. CPU servers managed 3 to 7. Fine-tuning is also a GPU job and isn't practical on a CPU VPS.

Yes, and neither one needs Docker. Running curl -fsSL https://ollama.com/install.sh | sh sets Ollama up as a systemd service on 127.0.0.1:11434, and you update it by running the script again. Ollama runs as its own ollama user and keeps models in /usr/share/ollama/.ollama/models. Open WebUI installs with pip install open-webui and starts with open-webui serve on port 8080. Set DATA_DIR so its data goes in a folder you choose, and run it as a systemd service.

No. Ollama has no login of its own, so anyone who reaches its port can use your models. In 2025 Cisco found 1,139 Ollama servers open to the internet, 214 of them serving models to anyone. Ollama binds to 127.0.0.1 by default, so leave OLLAMA_HOST alone and keep port 11434 closed. Put Open WebUI behind Apache with HTTPS, and expose only ports 80 and 443. Create the admin account right after install, since the first account to sign up becomes the admin. After that, open sign-up closes.

Open WebUI's data folder is the part to move, since it has webui.db with your users and chats, plus uploads and document embeddings. Stop Open WebUI and archive that folder. Restore it on the new server with the same WEBUI_SECRET_KEY, or everyone gets logged out. Models usually don't need copying. Run ollama pull on the new server to fetch them again, since each one is 2 to 10 GB. Copy any custom Modelfiles and your Ollama override.conf, which stores settings like context length and keep-alive.

RAM decides it, since the model has to fit with about 1 to 1.5 GB left for the operating system. At the default Q4_K_M setting, llama3.2:3b is 2.0 GB, llama3.1:8b is 4.9 GB and qwen3:14b is 9.3 GB. So the 4 GB plan suits 1B to 4B models for testing, and the 8 GB plan runs 7B and 8B models, which peak near 6.2 GB. The 16 GB plan fits gemma3:12b, qwen3:14b or an 8B model at Q8. Longer chats and several requests at once use more memory, and qwen3:30b-a3b, at 19 GB, won't fit on the 16 GB plan.

An 8B model needs the 8 GB plan, which is $19.99 a month on GreenGeeks VPS, renewing at $39.99. A CPU server isn't the cheap way to buy tokens, though. In 2026 testing, CPU servers worked out to about $4.41 to $9.04 per million output tokens when kept fully busy, and the cheapest API tiers charge far less. The server buys you privacy, a fixed bill with no rate limits, and a machine that can run other things too. Open models and Ollama cost nothing to download, so the plan is the main bill.

On a CPU VPS, about one person can chat comfortably at a time. In an August 2026 test, total output on CPU servers stayed near 7 tokens a second however many users were added. With 4 people at once, the slowest replies took about 128 seconds. Background work is different. An 8 vCPU server was estimated to handle jobs such as summaries, tagging and embeddings for about 15 to 18 users, where nobody is waiting on the reply. A single GPU card handled 16 requests at once in the same test.

Start with WEBUI_SECRET_KEY, because without the same key, a restored Open WebUI logs everyone out. Save it with Open WebUI's data folder, where the users, chats, uploads and document embeddings live. GreenGeeks leaves backups on a self-managed VPS to you, so nothing copies that folder unless you set it up. A nightly cron job can archive the folder and copy it to a second location off the server. Models can skip the backup, since ollama pull fetches them again.

GreenGeeks looks after the hardware and network. A self-managed VPS gives you full root access, so the operating system, Ollama and Open WebUI are yours to keep running. That means rerunning Ollama's install script to update it and upgrading Open WebUI with pip. When something breaks, journalctl -e -u ollama shows Ollama's logs, and ollama ps shows whether a model is loaded. Watch memory too. Switching models on an 8 GB server can run out of RAM, so set OLLAMA_MAX_LOADED_MODELS to 1.