CALC·04 · Local AI & GPUs
Local LLM VRAM & VPS vs Cloud GPU Cost Calculator
Calculate exact model weights + KV-cache memory for Ollama, OpenWebUI, and vLLM (Llama 3.3, DeepSeek-R1, Qwen 2.5) and decide between a KVM VPS or hourly Cloud GPU.

RATE SHEET
November 22, 2026Baseline Stack: Hostinger KVM (−20%) & Proton Privacy ($20 USD Credit)
| Provider / Tier | Hardware & Architecture | List | Net Benefit | Code | Link |
|---|---|---|---|---|---|
Hostinger KVM KVM 4 | 4 vCPU AMD EPYC · 16 GB DDR5 · 200 GB NVMe · 16 TB | $12.99/mo | ~$10.39/mo (-20%) | Claim -20% | |
Proton (VPN & Privacy) VPN Plus / Unlimited | WireGuard + NAT-PMP (Gluetun Docker) · Swiss No-Logs · Mail/Drive/Pass | Subscription | $20 USD credit | Claim $20 |
01 Open the link to bind XF3PLUTARXVF (-20%) or 5AA9X0DA ($20 USD).02 Pick 12/24-mo KVM VPS or Proton Plus/Unlimited.03 Deploy Ubuntu 24.04 + Docker + WireGuard tunnel.
INPUT · WEIGHTS & FP16 KV-CACHE MATH
Model, Quantization & Context Window
Quantization precision
16K tokens
2
4 h/day
ollama-cpu.shbash
# CPU/DDR5 KVM Deployment (Ollama + OpenWebUI) — est. 5–9 tok/s
ollama pull deepseek-r1:8b
OLLAMA_NUM_PARALLEL=2 OLLAMA_MAX_LOADED_MODELS=1 ollama run deepseek-r1:8bOUTPUT · VRAM FOOTPRINTCPU KVM VIABLE
Total
9.8 GB
Weights
4.6 GB
KV-Cache
4.00 GB
ARCHITECTURAL VERDICT
DeepSeek-R1 8B (9.8 GB) fits on KVM 4 (16 GB DDR5 at ~$10.39/mo with −20%) at ~5–9 tok/s for async n8n/private chat. For >60 tok/s use hourly GPU.
Engineering FAQs
Can you run Ollama on a CPU-only VPS without a GPU?
Yes for small 7B–14B Q4_K_M models (such as Qwen 2.5 7B or DeepSeek-R1 8B) on 16–32 GB DDR5 RAM (KVM 4 / KVM 8) at ~4–10 tokens/sec for async n8n batch jobs. For interactive multi-user chat or 32B–70B models, use hourly Cloud GPUs on hourly Cloud GPUs.