Skip to content

Solutions · Local models

A VPS for Ollama

Ollama serves open models over a local API, and the only hard rule is that the model must fit in memory: an 8-billion-parameter model is a 4.9 GB download, a 70B one is 42.5 GB. On a CPU it answers at a few tokens a second, which is enough for a private assistant, batch jobs and embeddings, and not enough for a chat product. BD-8 at $12.90 runs the 8B models; BD-64 at $84.90 runs the 70B ones.

Deploy BD-8Deploy C4-8-160 with Ollama installed

What it needs

Software from the docs, sizes labelled by source

The software and the security posture come from the project's own documentation. Where the project publishes no hardware figure, the sizes say whose they are.

From docs.ollama.com and the registry

  • Install: curl -fsSL https://ollama.com/install.sh | sh — creates the ollama user and a systemd service.
  • It listens on 127.0.0.1:11434 by default; OLLAMA_HOST=0.0.0.0 exposes it, and Ollama itself has no authentication, so put a proxy with auth in front first.
  • Models live in /usr/share/ollama/.ollama/models; sizes from the registry: llama3.1:8b 4.9 GB · qwen3:14b 9.3 GB · gemma3:27b 17.4 GB · qwen3:32b 20.2 GB · llama3.3:70b 42.5 GB.
  • A GPU is optional: without one the model runs on CPU and RAM, slowly; with one it runs from VRAM, fast.

Sizes (ours: the model must fit in RAM with headroom)

Sizing by workload
WorkloadvCPURAM
4B–8B models (gemma3:4b 3.3 GB, llama3.1:8b 4.9 GB)Assistants, summaries, embeddings. A few tokens a second on CPU.48 GB
12B–14B models (gemma3:12b 8.1 GB, qwen3:14b 9.3 GB)Noticeably better answers, noticeably slower.616 GB
27B–32B models (gemma3:27b 17.4 GB, qwen3:32b 20.2 GB)Usable for batch work, not for a waiting user.832 GB
70B models (llama3.3:70b 42.5 GB)Fits. Slow. A GPU server is the honest answer for interactive use.1664 GB

The plans that fit

4 machines, priced live

Ordered by memory, which is what decides the model. Prices are today's, per month, read from the catalogue.

  • BD-8$12.90/mo

    4 vCPU · 8 GB · 100 GB disk · 32 TB

    8 GB: the 8B models. 4 vCPU, 8 GB, 100 GB for downloads. Runs llama3.1:8b, qwen3:8b, gemma3:4b. Paste the one-line installer yourself.

    Deploy
  • U6-16-120$45.90/mo

    6 vCPU · 16 GB · 120 GB disk · Unmetered

    16 GB: the 14B models, unmetered, Istanbul. 6 vCPU and 16 GB run qwen3:14b and gemma3:12b; unmetered traffic for pulling models freely.

    Deploy
  • BD-64$84.90/mo

    16 vCPU · 64 GB · 500 GB disk · 32 TB

    64 GB: up to 70B, 16 cores. The cheapest machine on this page that fits llama3.3:70b or qwen3:32b in memory. Expect a couple of tokens a second on the big ones.

    Deploy
  • C4-8-160$72.90/mo

    4 vCPU · 8 GB · 160 GB disk · 4 TB

    8 GB, arrives with Ollama installed. The preset installs the service on a supplier that runs start-up scripts; 160 GB of disk for a library of models.

    Deploy installed

“Deploy installed” opens the configurator with our Ollama start-up script in the start-up box, in full, before you buy. It installs the service and pulls nothing; you choose the model. On the other rows the install is one pasted line.

The install

Four steps

Install, pull a model that fits, talk to it, keep it private.

  1. Install. curl -fsSL https://ollama.com/install.sh | sh The service starts on 127.0.0.1:11434 and survives reboots.
  2. Pull a model that fits this machine. ollama pull llama3.1:8b on 8 GB; qwen3:14b on 16 GB; gemma3:27b or qwen3:32b on 32 GB; llama3.3:70b on 64 GB.
  3. Use it. ollama run llama3.1:8b, or the API: curl http://127.0.0.1:11434/api/generate -d '{"model":"llama3.1:8b","prompt":"hello"}'. From your laptop, ssh -N -L 11434:127.0.0.1:11434 root@your-server.
  4. Only then expose it, behind auth. systemctl edit ollama.service with Environment="OLLAMA_HOST=0.0.0.0:11434" makes it reachable; a reverse proxy with a password or an API key in front is not optional, because Ollama has none.

Honestly

What we would actually buy

Buy for the model you will actually run, and know what CPU inference feels like before you do.

For the 8B models

BD-8 · $12.90/mo

4 vCPU and 8 GB at $12.90 run llama3.1:8b or qwen3:8b with room to spare, at a few tokens a second: right for a private assistant you talk to yourself, nightly summaries, or embeddings for a search index. If you want 27B-class answers, BD-64 at $84.90 holds them and the 70B models too, slowly. If people will wait on the answers, stop reading this page and look at the GPU line.

Deploy BD-8BD-64 for the big models

Questions

What people ask before they buy one.

How much RAM does Ollama need?
As much as the model plus a little headroom: the model file is loaded into memory. Ollama's registry lists llama3.1:8b at 4.9 GB, qwen3:14b at 9.3 GB, gemma3:27b at 17.4 GB, qwen3:32b at 20.2 GB and llama3.3:70b at 42.5 GB, so the tiers are 8, 16, 32 and 64 GB. The operating system and Ollama itself want another gigabyte or two. A model that does not fit either fails to load or swaps, which is worse.
How fast is it on a CPU VPS?
Slow, and honestly so: a few tokens a second for an 8B model on four cores, less for bigger ones. That is fine for jobs that run unattended and for one person chatting, and it is not fine for a product with users waiting. More cores help a little; memory bandwidth is the real limit. For interactive speed you want a GPU, which is a different page.
Do I need a GPU?
Not to run Ollama. You need one to run it fast. Everything on this page is CPU inference; the GPU line has cards from a single L4 upward, and a 7B model that crawls on CPU answers instantly there. Decide by who is waiting for the answer.
What does the start-up preset do?
It runs Ollama's official installer, which creates the ollama user and a systemd service bound to 127.0.0.1:11434, enables it, and opens nothing but SSH on the firewall. It pulls no model — that choice depends on the machine you bought — and leaves /root/OLLAMA-NEXT-STEPS.txt with the pull commands per RAM tier, the tunnel command and the warning about exposing the API.
Can other machines use it?
Yes, once you set OLLAMA_HOST=0.0.0.0:11434 through systemctl edit ollama.service and open the port. Do not do that without a reverse proxy that requires a key or a password: Ollama has no authentication of its own, and an open Ollama is a free GPU-less compute donation to the internet. Over SSH tunnels or a VPN, none of this is needed.
Where do the models go, and how big do they get?
/usr/share/ollama/.ollama/models by default, changeable with OLLAMA_MODELS (the ollama user must own the directory). A modest library — an 8B, a 14B and a 32B — is about 35 GB, which is why every row on this page has 100 GB or more.

Keep reading

Related