Why Run an LLM on a VPS at All
If you are shopping for the best vps for hosting llm experiments, you are probably chasing three things: privacy, predictable cost, and a sandbox where open models like Llama and Mistral run on your terms. A virtual private server gives you full root access, no per-token API billing, and prompts that never leave your own machine.
Realistic uses include a private chat interface with Open WebUI, retrieval-augmented generation prototypes, small-model fine-tuning tests, and a local-style API endpoint for side projects. The honest caveat comes early: a CPU-only VPS generates tokens slowly, so think experiments and learning, not production chatbots.
Hostinger is a practical starting point because it ships an Ollama VPS template with Llama 3 and Open WebUI pre-installed, and its VPS hosting plans scale from 4 GB to 32 GB of RAM. That covers everything from first experiments to serious 13B tinkering, checked 13 Oct 2026.
How Much Memory Each Model Size Really Needs
Memory is the whole game, and the ballparks below come from current hardware guides, checked 13 Oct 2026. All assume 4-bit quantization (GGUF Q4_K_M), which shrinks weights to roughly half a byte per parameter with modest quality loss. A handy rule of thumb: working memory in GB is about half the parameter count.
- 7B-8B models (Llama 3.1 8B, Mistral 7B): the model file is around 4.5 to 5 GB; budget about 8 GB of RAM including context and the OS.
- 13B-14B models (Qwen 2.5 14B): the file is roughly 8 to 10 GB; budget about 16 GB of RAM.
- 30B-32B models: the file lands near 20 GB; budget 24 to 32 GB of working memory.
- 70B models (Llama 3.1 70B): the file alone is about 43 GB; working memory runs 48 to 52 GB with context.
One more trap: the KV cache grows with context length, so long conversations and agentic loops eat memory fast. CPU-only inference is usable at 7B-8B, patience-testing at 13B, and pointless beyond that.
Ollama vs vLLM vs llama.cpp on a CPU VPS
Ollama is the easiest path: one install script, automatic GGUF downloads, and an OpenAI-compatible API on port 11434. Hostinger’s own tutorial at hostinger.com/tutorials/openclaw-ollama/ walks through installing it on a Hostinger VPS, and the pre-built Ollama template removes setup entirely.
vLLM delivers the highest serving throughput, but it is GPU-only (CUDA), which makes it pointless on a CPU-only VPS. Skip it unless you move to GPU hardware.
llama.cpp gives maximum control over threads, GPU offloading, and quantization settings, and it is the engine under Ollama’s hood. Start with Ollama; reach for llama.cpp directly only when you need to squeeze performance from weak CPUs. For related self-hosted AI setups, see Hostinger KVM 2 pricing and our n8n VPS guide.
The Best VPS for Hosting LLM Workloads on Hostinger
Map the model to the plan and the choice gets simple. KVM 1 (1 vCPU, 4 GB RAM, 50 GB NVMe) can technically hold a 7B model, but the OS leaves almost no headroom. KVM 2 (2 vCPU, 8 GB RAM, 100 GB NVMe) is the sweet spot for 7B-8B experiments. KVM 4 (4 vCPU, 16 GB RAM, 200 GB NVMe) is the realistic minimum for 13B models. KVM 8 (8 vCPU, 32 GB RAM, 400 GB NVMe) tempts for 32B, but CPU inference at that size is glacial.
Two hard limits, checked 13 Oct 2026: Hostinger VPS plans are CPU-only with no GPU option, and 70B is not viable on any of them. Disk space is generous by comparison; even the 43 GB 70B file fits on the smallest plan, so memory, not storage, is the wall. Compare plans on our VPS hosting overview, and see self-hosting Supabase on a VPS for another self-hosted workload.
Renewal Reality and Who This Setup Is For
Renewal reality first: Hostinger's deep intro discounts expire, and renewals land roughly 3.3 to 4.3 times the introductory rate according to 2026 reporting, so budget for the renewal figure shown at checkout, not the headline price. The 30-day money-back guarantee softens the first term only. Read our renewal price breakdown before committing to a long term.
Who this is for: developers prototyping with open models, privacy-conscious teams keeping data in-house, students learning LLM operations, and anyone who wants a cheap always-on inference sandbox. VPS plans are self-managed, so OS patching, firewall rules, and Ollama updates are your job. If you need a managed GPU or an SLA, this is not your setup.
When NOT to Pick a CPU VPS for LLMs
Be blunt with yourself: anything involving 70B models, fine-tuning, production inference, or latency-sensitive apps does not belong on a CPU VPS. For that work, a GPU cloud is honestly better. RunPod and Lambda Labs rent GPUs like the RTX 4090 by the hour (RunPod listed around $0.70 per hour in September 2026 pricing), and you pay only while the instance runs.
Also consider whether you need to self-host at all. If you just want answers from a capable model, a hosted API is cheaper and faster than any VPS experiment. A budget VPS is a learning lab and a privacy sandbox, not a production inference platform.
| Model size (4-bit GGUF) | Memory needed | Realistic on | Hostinger VPS fit |
|---|---|---|---|
| 7B-8B (Llama 3.1 8B, Mistral 7B) | About 8 GB RAM with context | CPU VPS or small GPU | KVM 2 is the sweet spot; KVM 1 works but is tight |
| 13B-14B (Qwen 2.5 14B) | About 16 GB RAM with context | CPU VPS at low speed, or 16 GB GPU | KVM 4 fits; expect slow CPU inference |
| 30B-32B class | About 24-32 GB RAM | 24 GB GPU or big-memory CPU box | KVM 8 only; painfully slow without a GPU |
| 70B (Llama 3.1 70B) | About 48-52 GB with context | Dual GPU or 96 GB+ unified memory | Not viable on any Hostinger CPU VPS |
VPS KVM 2
24-month plan, paid upfront
Renews at $14.99/mo
- 2 vCPU + 8 GB RAM
- 100 GB NVMe storage
- 8 TB bandwidth
- Full root access
- n8n one-click template available
FAQ
Yes. The 4-bit quantized file is under 5 GB, so an 8 GB VPS holds the model plus working context with room to spare. Expect a few tokens per second on CPU, which is fine for experiments and private chat but not for anything latency-sensitive. Give the model a full core or two and keep context windows modest for the best experience.
Around 16 GB. The 13B Q4 file is roughly 8 to 10 GB, and the KV cache plus operating system overhead pushes working memory to about 14 to 16 GB. Hostinger KVM 4 with 16 GB RAM is the realistic minimum; anything smaller will swap and crawl. CPU inference at 13B is slow, so treat it as a learning setup rather than a production one.
No. A 70B model at 4-bit quantization needs roughly 48 to 52 GB of memory with working context, while KVM 8 tops out at 32 GB RAM and has no GPU. You would need dual GPUs, a 96 GB unified-memory machine, or a GPU cloud like RunPod. The 43 GB model file would fit on the 400 GB disk, but disk space is not the bottleneck here.
Ollama. It installs in one command, downloads GGUF models automatically, and runs on CPU or GPU, and Hostinger even ships an Ollama VPS template with Open WebUI. vLLM delivers higher throughput but requires an NVIDIA GPU, which Hostinger VPS plans do not include. Use llama.cpp directly only if you want maximum low-level control over threading and quantization.
No. As of 13 Oct 2026, Hostinger VPS plans are CPU-only KVM instances ranging from 1 vCPU with 4 GB RAM up to 8 vCPU with 32 GB RAM, with no GPU add-on. They suit small quantized models and inference experiments. For GPU inference or fine-tuning, rent from a GPU cloud such as RunPod or Lambda Labs instead of forcing the workload onto a CPU VPS.
Intro discounts are steep and renewals land much higher, roughly 3.3 to 4.3 times the introductory rate according to 2026 reporting. Check the renewal figure on the pricing page before you commit, factor it into your project budget, and remember the 30-day money-back guarantee covers your first term only. Longer terms lock in the intro rate for longer.