Can You Run Ollama on a 2GB VPS? Only If You Want a 1B Model

Yes, technically — with a 1B model. Ollama itself is small; what doesn’t fit is the model.

And the arithmetic settles this before speed ever comes into it. If the model file is larger than your available RAM, it doesn’t run. Tokens per second is a question you only get to ask once it loads.

What fits in 2GB

A model’s file size is roughly its parameter count multiplied by the bytes per parameter, which quantisation determines. At Q4 — the common default — that’s a bit over half a byte per parameter.

Now subtract the operating system. A 2GB VPS running Linux plus the Ollama runtime leaves roughly 1.4 to 1.6 GB for the model — and you need headroom above the file itself for the KV cache that holds your conversation context.

So 2GB gives you a 1B or 1.5B model and nothing above it. Not “slowly” — at all. The 3B is already over the line.

What a 1B model is actually good for

This is the part worth being straight about, because “it runs” and “it’s useful” are different claims.

A 1B model is genuinely capable of: simple classification, short summarisation, basic text cleanup, keyword extraction, and structured-output tasks where the format matters more than the reasoning.

It is not capable of: the coding help, multi-step reasoning or long-context work that people typically install Ollama for. If you tried a 7B or 8B model elsewhere and want that on a 2GB box, the answer is no — and no amount of quantisation closes a 2.5× gap.

And on CPU-only inference generally: expect single-digit tokens per second on small models, which is usable for a background job and frustrating for anything interactive. We have not benchmarked this — see the methodology — but the direction is not controversial.

The three ways out, costed

If the 1B ceiling doesn’t work, there are exactly three routes. Here’s what each costs monthly.

1. A bigger CPU box

8GB is the comfortable floor for a 7–8B model at Q4, because you need the 3.9 GB file plus context plus OS.

Slow but always available, and predictable. Note how far apart those are — Hetzner is no longer the cheap option since its 2026 repricing.

2. Rent a GPU by the hour

The break-even against an 8GB CPU box is about 44 hours a month — roughly 1.5 hours a day. Below that, hourly GPU is both cheaper and dramatically faster. Above it, the CPU box wins on cost and loses on speed.

This is the option most people should take and don’t, because intermittent use is the actual usage pattern for a personal model.

3. Use an API and stop paying for idle hardware

$2–15/month for typical personal use, from our cost research — and no server to maintain, patch or size.

The honest comparison: a hosted API is usually cheaper than any of the above and always faster. The reasons to self-host are privacy, control, and not sending your data to a third party — which are real reasons, but they’re not cost reasons. We ran that arithmetic properly and the API wins on price by a wide margin.

So what should you do

If you have a 2GB box already: install Ollama, pull a 1B model, and use it for the narrow tasks it handles. It costs nothing to try and the ceiling is clear.

If you’re buying hardware to run a 7B model: buy 8GB, not 2GB, and expect single-digit tokens per second on CPU.

If you want the model to feel responsive: rent a GPU by the hour and only pay while you use it.

If you just want the capability: use an API. The self-hosting decision should be about data, not about saving money — because it usually doesn’t.

The providers

Hetzner — German company, German data centres, and the jurisdiction answer if that’s why you’re self-hosting in the first place. But check the price before assuming it’s cheap: after the 2026 repricing its 8GB CPX32 is ~$38.87, well above Hostinger, and its cost-optimized tier has been showing as unavailable.

Hostinger — the cheapest way to get to 8GB at $14.99, with one-click Docker templates that include Ollama and remove the install step where most first attempts fail. If you’re stepping up from 2GB specifically to run a real model, this is the direct answer.

RunPod — hourly GPU from $0.34 on its community tier, billed per second. The right shape for intermittent use, and the option that turns “unusably slow” into “fast” for less money than a permanent 8GB box if you use it under 1.5 hours a day.

DigitalOcean — $48 for 8GB, the most expensive here, bought for documentation quality and predictable APIs rather than price.

For the fuller sizing picture, how much RAM Ollama needs per model covers the tiers above 2GB, and Ollama VPS cost has the storage and idle-billing details that catch people out.

How we checked this

The memory table is arithmetic, and it’s the load-bearing part of this article. Model file size is approximately the parameter count multiplied by bytes per parameter, which quantisation sets — roughly 0.55 bytes at Q4_K_M, 0.68 at Q5_K_M, 1.06 at Q8_0 and 2.0 at FP16. Those per-parameter figures are the widely-used approximations for GGUF quantisation formats, not measurements we took, and real files vary by a few percent with architecture and tokeniser size. The conclusion is robust to that variance — a 7B at Q4 is around 4 GB whichever way you round it, and a 2GB box has about 1.5 GB free.

The OS overhead figure of roughly 0.4–0.6 GB is a conventional estimate for a minimal Linux install plus the Ollama runtime, not something we measured on a 2GB VPS.

We did not benchmark anything. The brief for this article asked for measured tokens per second across CPU-only VPS tiers, and we haven’t run those tests — we have no representative VPS fleet, and CPU inference throughput varies enough with core count, memory bandwidth and neighbours that a single figure would mislead. The “single-digit tokens per second” characterisation is the widely-reported range, not our measurement, and we’ve flagged it as such in the body rather than presenting a table of numbers we didn’t produce.

The cost fork uses figures from our own research across this series in August 2026: Hostinger’s 8GB at $14.99 and DigitalOcean’s at $48 as published renewal and flat rates, Hetzner’s CPX32 at €35.99 converted at approximately 1.08, RunPod’s community-tier RTX 4090 at $0.34/hour, and the $2–15 monthly API range from our self-hosted-versus-API cost analysis. The break-even of ~44 hours per month is our arithmetic on the Hostinger figure and would shift with a different CPU box.

What we did not do: run Ollama on a 2GB VPS, pull any model, measure load times or throughput, or verify Ollama’s own published minimums against these numbers.

The host links above are affiliate links. This article’s conclusion is that a hosted API is usually cheaper and faster than any of them, and that self-hosting should be a data decision rather than a cost one.

FAQ

Can Ollama run on a 2GB VPS?

Yes, with a 1B or 1.5B model at Q4 quantisation — around 0.6–0.8 GB. Anything larger doesn’t fit: a 3B needs 1.7 GB and a 2GB box has roughly 1.5 GB free after the OS.

Can I run a 7B model on 2GB of RAM?

No. A 7B model at Q4 is about 3.9 GB, more than two and a half times the available memory. Quantising further degrades quality long before it closes that gap.

How much RAM do I need for a 7B model?

8GB is the comfortable floor — the ~3.9 GB file plus context and operating system. That’s the tier to buy if a 7–8B model is the goal.

Is a 1B model useful for anything?

Yes, for narrow tasks: classification, short summarisation, text cleanup, keyword extraction, structured output. Not for coding help or multi-step reasoning.

Is it cheaper to rent a GPU than buy a bigger VPS?

Below about 44 hours a month — roughly 1.5 hours a day — yes, and it’s far faster. Above that, a permanent 8GB CPU box costs less.

How fast is Ollama on a CPU-only VPS?

Slow — single digit tokens per second on small models is the commonly reported range. We haven’t benchmarked it, and it varies with cores, memory bandwidth and what else is on the host.

Should I self-host a model to save money?

Usually not. A hosted API runs $2–15/month for typical personal use, which beats a permanent server. Self-host for privacy and control, not for cost.