Skip to content

Best VPS for Ollama in 2026

Updated Aug 2026 · 8 plans benchmarked

Ollama hosting is a RAM story: the model weights must fit in memory or nothing runs, so the winning box is the one with the most GB per dollar. CPU-only inference works fine — no GPU required — but tokens per second is what you'll feel, especially if agents or automations sit on the API, so fast cores and headroom above the model size matter.

Prefer to try before you commit? Kamatera's free tier lets you spin one up now.Start free on Kamatera →
What to look for
  • Size RAM to the model: 8 GB runs 3B models comfortably and 7B with tight context, 16 GB is the 7B sweet spot, 32 GB unlocks 13B, and 30B+ models want 64 GB or more.
  • Be honest about CPU-only: cheap VPS plans have no GPU, and a 7B model on a 4-core box generates roughly 5–10 tokens per second — fine for chat and background automation, too slow for real-time generation at scale.
  • Budget disk for weights — each model is roughly 5–20 GB, and NVMe keeps pulls and model loads from stalling.
  • Never expose port 11434 to the internet: Ollama's API has no authentication, so bind it to localhost and put a TLS reverse proxy (Caddy or Nginx) in front.

The picks

ContaboBest overall

The most RAM per dollar on this list — the entry Cloud VPS 4 puts 8 GB and 100 GB of NVMe under the lowest price here, and the 24 GB Cloud VPS 8 reaches 13B-class models for less than most rivals charge for 8 GB.

From $4.95/mo · 4 vCPU / 8 GB / 100 GB · EU + US + UK + Asia + AustraliaGet Contabo
OVHcloudEasiest to start

The VPS-2 (4 vCPU / 8 GB) covers small models with unlimited traffic, anti-DDoS, and daily backups included, and the range steps to 24 GB without leaving NVMe.

From $4.54/mo · 2 vCPU / 4 GB / 40 GB · US + EU + Canada + Asia + AustraliaGet OVHcloud
Hetzner CloudRunner-up

The CX32 (4 vCPU / 8 GB NVMe) clears the small-model floor, and the CX42 (8 vCPU / 16 GB) is the cheapest fast-core route here to the 16 GB that 7B models really want.

From $4.59/mo · 2 vCPU / 4 GB / 40 GB · EU + USGet Hetzner Cloud
affiliate disclosure

The picks, side by side

ContaboBest overall
From$4.95/movCPU4RAM8 GBStorage100 GBRegionsEU, US, UK, Asia, Australia
OVHcloudEasiest to start
From$4.54/movCPU2RAM4 GBStorage40 GBRegionsUS, EU, Canada, Asia, Australia
Hetzner CloudRunner-up
From$4.59/movCPU2RAM4 GBStorage40 GBRegionsEU, US

Prices verified Aug 2026 · How we benchmark →

Common questions

How much RAM does Ollama need?

Ollama itself is light — the model is what needs memory. 8 GB runs 3B models comfortably and 7B models with tight context; 16 GB is the sweet spot for 7B with reasonable context; 13B wants 32 GB and 30B+ wants 64 GB or more.

Can Ollama run on a CPU-only VPS?

Yes — Ollama falls back to CPU inference automatically, no GPU needed. Expect roughly 5–10 tokens per second for a 7B model on a 4-core box: fine for chat, summarization, and automation, but noticeably slower than the hosted APIs you may be used to.

Do I need a GPU VPS for Ollama?

Only for speed or scale. A GPU delivers 5–10x the tokens per second but multiplies the price, and the same sizing logic applies — a 7B model still needs ~4–6 GB of VRAM. For a personal or small-team setup, a well-sized CPU box is usually the better value.

How much disk space do Ollama models take?

Roughly 2–5 GB for 3B models, 4–8 GB for 7B, and 20 GB+ for the largest you'd run on a VPS. Weights are re-downloadable, so disk is about convenience, not backup — but keep enough free space for two or three models at once.

Search SelfHost Atlas

Search apps, comparisons, guides, and categories.

We use analytics cookies (Google Analytics, PostHog) to see which guides are useful. No ad networks, no cross-site tracking. See our privacy policy.