This VPS RAM calculator tells you how much memory your self-hosted stack needs and which of the servers we tested is the cheapest fit. Every app figure in it comes from our Lab. For example, n8n idled at 379.7 MiB on our Hetzner CX23, and Llama 3.2 3B added up to 2,945 MiB while loaded.
RAM decides more than CPU for this kind of stack. When an app runs out of memory it does not slow down, it stops: Llama 3.1 8B refused to load on our 4 GB CAX11 with "insufficient memory". It also costs money: in Hetzner's 2026 prices, going from 4 GB to 8 GB (CX23 to CX33) moves you from 6.64 € to 10.27 € a month. More calculators are in our tools section.
Tested in the Lab Hetzner CX23, CX33, CX43, CAX11, CAX21 · Nov 2026
How it works
The calculator adds up a fixed base and one block per app you tick, then picks the cheapest tested server whose RAM, minus 15% headroom, is at least that total. All figures are in MiB (mebibytes; 1,024 MiB = 1 GiB, which the result shows as GB).
| Block | RAM it adds | Where the number comes from |
|---|---|---|
| Ubuntu + Docker | 550 MiB | 400 MiB for the OS plus 150 MiB for Docker. Our fresh servers used 308 to 500 MiB; with Docker installed, 359 to 669 MiB. |
| n8n + Postgres | 450 + 70 MiB idle, plus 400 MiB headroom | n8n 2.46.6 idled at 379.7 MiB on our CX23 and at 414.9 and 416.6 MiB on the CX33 and CX43. Postgres 16 idled at 69.6 MiB. The 400 MiB covers workflow runs: n8n used 785.3 MiB right after start. |
| Supabase | 1,562 MiB | The highest container total we measured for the official 11-container stack after 90 s idle (on the CAX11). Other servers: 1,101 to 1,399 MiB. |
| Llama 3.2 1B | 1,978 MiB | System RAM used while loaded minus fresh-server RAM, highest across servers (CAX21: 2,315 − 337). |
| Llama 3.2 3B | 2,945 MiB | Same method (CAX21: 3,282 − 337). |
| Llama 3.1 8B | 5,523 MiB | Same method (CX43: 6,023 − 500). It did not run on the 4 GB CAX11. |
| Each extra small app | 150 MiB | An assumption for a bot or small site, not a measurement. |
Using the highest value across our servers keeps the estimate on the safe side. The model figures are also slightly generous, because the "fresh" baseline was taken before Docker was installed.
The measurements behind it
Supabase, all five servers:
| Server | Containers | RAM (containers) | RAM used (system) | Fits? | Image pull | Disk (images) |
|---|---|---|---|---|---|---|
| Hetzner CX232 vCPU · 4 GB · x86_64 | 11 | 1,399 MiB | 1,594 MiB | yes | 111 s | 9.563GB |
| Hetzner CX334 vCPU · 8 GB · x86_64 | 11 | 1,316 MiB | 1,732 MiB | yes | 50 s | 9.563GB |
| Hetzner CX438 vCPU · 15 GB · x86_64 | 11 | 1,101 MiB | 1,870 MiB | yes | 29 s | 9.563GB |
| Hetzner CAX112 vCPU · 4 GB · aarch64 | 11 | 1,562 MiB | 1,564 MiB | yes | 99 s | 10.14GB |
| Hetzner CAX214 vCPU · 8 GB · aarch64 | 11 | 1,318 MiB | 1,712 MiB | yes | 52 s | 10.14GB |
Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data
n8n install on our production CX23:
| n8n 2.46.6 on Hetzner CX23 | Measured |
|---|---|
| Docker install | 22 s |
| Image pull + first start to healthy | 105 s |
| n8n RAM right after start | 785.3 MiB |
| n8n RAM idle (after ~60 s) | 379.7 MiB |
| PostgreSQL 16 RAM idle | 69.6 MiB |
| Whole server RAM used, idle | 1036 MiB of 3,814 |
| n8n image size | 1.94GB |
| Disk used after install | 4.7G of 38G |
Measured by AI Hosting Lab on 2026-11-05 · docker compose: n8n + postgres:16, docker.io from Ubuntu repo, task runners on · method
Ollama models, including system RAM while loaded:
| Server | Model | Generation tokens/s | Prompt tokens/s | Load time | RAM used (system) | Note |
|---|---|---|---|---|---|---|
| Hetzner CX232 vCPU · 4 GB · x86_64 | Llama 3.2 1B | 10.3 | 454.0 | 3.5 s | 2,001 MiB | |
| Hetzner CX232 vCPU · 4 GB · x86_64 | Llama 3.2 3B | 5.8 | 325.2 | 8.8 s | 3,002 MiB | |
| Hetzner CX232 vCPU · 4 GB · x86_64 | Llama 3.1 8B | did not run | – | – | – | OOM-killed after ~50 s: Ollama fit check reported it would leave 1698 MiB free, |
| Hetzner CX334 vCPU · 8 GB · x86_64 | Llama 3.2 1B | 23.2 | 1,322.8 | 2.0 s | 2,205 MiB | |
| Hetzner CX334 vCPU · 8 GB · x86_64 | Llama 3.2 3B | 13.5 | 776.5 | 3.7 s | 3,068 MiB | |
| Hetzner CX334 vCPU · 8 GB · x86_64 | Llama 3.1 8B | 6.3 | 258.6 | 10.0 s | 5,941 MiB | |
| Hetzner CX438 vCPU · 15 GB · x86_64 | Llama 3.2 1B | 37.5 | 1,910.9 | 1.7 s | 2,293 MiB | |
| Hetzner CX438 vCPU · 15 GB · x86_64 | Llama 3.2 3B | 21.2 | 1,161.3 | 3.7 s | 3,196 MiB | |
| Hetzner CX438 vCPU · 15 GB · x86_64 | Llama 3.1 8B | 9.7 | 447.4 | 7.9 s | 6,023 MiB | |
| Hetzner CAX112 vCPU · 4 GB · aarch64 | Llama 3.2 1B | 15.0 | 835.5 | 4.8 s | 2,203 MiB | |
| Hetzner CAX112 vCPU · 4 GB · aarch64 | Llama 3.2 3B | 6.5 | 376.4 | 6.0 s | 3,214 MiB | |
| Hetzner CAX112 vCPU · 4 GB · aarch64 | Llama 3.1 8B | did not run | – | – | – | llama-server process has terminated: exit status 1: ggml_aligned_malloc: insuffi |
| Hetzner CAX214 vCPU · 8 GB · aarch64 | Llama 3.2 1B | 26.8 | 1,427.1 | 3.2 s | 2,315 MiB | |
| Hetzner CAX214 vCPU · 8 GB · aarch64 | Llama 3.2 3B | 12.2 | 664.1 | 5.2 s | 3,282 MiB | |
| Hetzner CAX214 vCPU · 8 GB · aarch64 | Llama 3.1 8B | 5.7 | 226.5 | 12.3 s | 5,796 MiB |
Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data
What typical stacks come out at
Results from the calculator with its current data:
- n8n only: about 1.4 GB. Cheapest fit: CX23 (4 GB). More in n8n requirements.
- n8n + Supabase: about 3.0 GB. The CX23 fits, but with little room; see our Supabase VPS sizing.
- n8n + Llama 3.2 3B: about 4.3 GB. Cheapest fit: CX33 (8 GB).
- n8n + Llama 3.1 8B: about 6.8 GB. This goes past 85% of an 8 GB server, so the calculator picks the CX43 (16 GB).
- n8n + Supabase + Llama 3.1 8B: about 8.4 GB. Cheapest fit: CX43.
Assumptions and limits
- Idle, not load. n8n and Supabase were measured idle. Big workflows, a large database or many users use more. Under our load test n8n peaked at 1003.0 MiB on the CX23 and 1049.6 MiB on the CX33 at 50 concurrent requests, so the calculator adds 400 MiB of headroom for n8n.
- Short model answers. Ollama figures come from a 256-token answer with the default context. Ollama's FAQ says required RAM scales with the number of parallel requests times the context length, so longer contexts and parallel requests need more. We have not measured those yet.
- No swap. Our test servers had no swap configured (0 MiB). Swap can keep a server alive in a pinch, but a model swapping to disk becomes very slow.
- Only servers we tested. The recommendation picks from five Hetzner plans: CX23, CX33, CX43, CAX11 and CAX21. If nothing fits, it tells you to look at 16 GB+ plans. Other providers are not yet tested in our Lab; testing is scheduled.
- Prices. The price shown is the Hetzner gross monthly price (including VAT) from the Hetzner Cloud API, checked 5 November 2026.
- Linux counts RAM its own way. Our "system RAM used" comes from
free -m, and container RAM fromdocker stats --no-stream, which on Linux subtracts cache from a container's memory use. Monitoring tools that count cache will show higher numbers for the same server.
Next step: if a local model is on your list, see what its speed is like on each server in VPS for Ollama.
Hetzner prices on this page include 21% VAT, as charged to our Czech (EU consumer) account by the Hetzner API on 2026-11-05. Without VAT (businesses, most buyers outside the EU): CX23 5.49 €, CX33 8.49 €, CX43 15.99 €, CAX11 5.99 €, CAX21 10.49 € per month.
FAQ
How much RAM does n8n need on a VPS?
On our Hetzner CX23, n8n 2.46.6 idled at 379.7 MiB and Postgres 16 at 69.6 MiB, with 1,036 MiB used on the whole server. The calculator adds 400 MiB of headroom for workflow runs, so n8n alone fits a 4 GB server easily.
Can I run n8n and Supabase on a 4 GB VPS?
The calculator says about 3.0 GB, which fits a 4 GB server with 15% spare. Both numbers are idle measurements, so for production we would still pick 8 GB.
How much RAM does a local LLM need?
While loaded, Llama 3.2 3B added up to 2,945 MiB and Llama 3.1 8B up to 5,523 MiB to the system RAM in our tests. Llama 3.1 8B did not run on a 4 GB server at all.
Is the calculator accurate for my traffic?
It is built from idle measurements plus fixed headroom. Heavy workflows, big databases, many users or long model contexts need more. Treat the result as the smallest size worth trying.