Disclosure: this page contains affiliate links. If you buy through them we may earn a commission at no extra cost to you.

Budget VPS plans don’t include GPUs, so serious model inference means CPU — and that means RAM is your real constraint. Here’s how to size it.

RAM is the bottleneck

  • Small models (7B quantized, ~4-bit) — 8 GB RAM minimum.
  • Embedding or classifier models — 2–4 GB is usually fine.
  • LLM fine-tuning — not realistic on a VPS; use a cloud GPU.

The pick: Hostinger KVM 2

Hostinger’s KVM 2 (8 GB RAM, ~$8.99/mo intro) is the entry point for running a quantized 7B model with CPU inference. Faster NVMe storage helps model loading. The KVM 1 (4 GB, ~$6.49/mo) works for embeddings and small classifiers.

Make CPU inference usable

./llama-server -m model-q4.gguf -c 2048
  • Quantize to 4-bit to cut memory roughly in half.
  • Expect slower tokens/second than a GPU — fine for batch or low-traffic use.
  • Offload heavy jobs to a GPU cloud when you need speed.

Bottom line: embeddings and classifiers → 4 GB; quantized 7B LLM → 8 GB. CPU-only means patience, not impossible.

Check Hostinger's VPS plans