Ollama on a CPU VPS: which models fit in 16 to 64 GB of RAM

Ollama runs language models on a VPS without GPU, but on a processor the answers stay slow. With 16 GB of RAM, small quantized models up to about 8 billion parameters are realistic; 32 GB opens intermediate models, and 64 GB lets you try very large ones, with patience. The test to run: install Ollama, load a model, then measure the RAM really used.

• On CPU, a VPS runs small quantized models, with slow answers.
• 16 GB is enough for about 8 billion parameters; 64 GB opens the larger ones, slowly.
• Ollama listens locally by default: do not open it to the internet without protection.
• The Performance plan gives 48 GB of RAM at 21.99 € as of 7 November 2026.

Install Ollama

On Debian or Ubuntu, Ollama installs with its official script: curl -fsSL https://ollama.com/install.sh | sh. Read the script before running it if you wish, then check the installation with ollama --version.

• The service listens by default on 127.0.0.1, port 11434: it is not reachable from the internet.
• Without GPU, Ollama uses the processor and the RAM of the VPS.

Choose a model according to RAM

A 4-bit quantized model takes roughly half its number of billions of parameters in gigabytes, plus memory for the context and the system. These are rough figures: the exact size is shown on the model page in the Ollama library.

• 16 GB: models up to about 8 billion parameters, keeping a margin for the system.
• 32 GB: models of about 14 to 32 billion parameters, slower.
• 48 to 64 GB: models of about 70 billion parameters, for uses where slowness is acceptable.

Run and measure

Download a model with ollama pull followed by its exact name, for example qwen2.5:7b, then run it with ollama run. During generation, ollama ps shows the size loaded in memory and free -h the RAM left.

• Stop the other services during the measure, for a reliable figure.
• If the system starts using swap, choose a smaller model.

Do not expose the port

By default, Ollama is only reachable from the server. If you set OLLAMA_HOST to open it to the network, protect it: firewall, restricted access or SSH tunnel. An open port 11434 on the internet can be used by anyone, at your expense in computing and memory.

• For personal use, an SSH tunnel is enough: ssh -L 11434:localhost:11434 user@vps.
• See secure a Linux VPS for the firewall.

Why choose ACLClouds for local AI

VPS plans are described in vCPU and GB of RAM, the right criteria for a model run on a processor. The Performance plan gives 48 GB for 21.99 € as of 7 November 2026, and the Pro plan 64 GB for 34.99 €.

• The Essentiel plan, 16 GB of RAM at 9.99 €, covers small models.
• KVM virtualization and servers located in France.
• Control panel available in 7 languages.