Self-hosted LLM on a budget VPS: Ollama and 7B–13B in 2026
Run a self-hosted LLM on a budget VPS in 2026 with Ollama and llama.cpp. Sizing, quantisation and privacy for 7B–13B models under $20/month.
Self-hosting an LLM on a budget VPS stopped being a hobbyist hack in 2026. With Q4 quantisation, a 7B model fits under 5 GB of RAM and a comfortable 13B fits under 10 GB — both within reach of a small VPS that invoices in crypto and never asks for identity.
How much RAM a quantised model really needs
- ▹1–3B (Q4): ~2–4 GB — fits the smallest instances, fast CPU inference.
- ▹7B (Q4_K_M): ~4.5–6 GB — the sweet spot for a single-user server.
- ▹13B (Q4_K_M): ~8–10 GB — good quality, slower tokens on CPU.
- ▹70B (Q4): ~40 GB — dedicated hardware territory, not budget VPS.
Ollama vs llama.cpp on a CPU server
Both runtimes expose an OpenAI-compatible endpoint. Ollama is the fastest path from install to a working local API — one service on port 11434, models pulled by name. llama.cpp gives you lower-level control over thread count, context size and memory compression when you need to squeeze a slower box.
A private endpoint that stays yours
The privacy argument is structural: prompts and responses never leave the server. Keep the model endpoint bound to 127.0.0.1 and reach it over your own WireGuard tunnel instead of exposing it publicly — the same pattern you already use for a Monero node or an RPC endpoint.