🟢 All systems operational · MIT-licensed runner on hardware you rent
+ Add Provider
Open source · MIT licence

Ollama — One Command,
Then It’s Your Model

Ollama pulls a model and puts it behind an HTTP API in a single command. It is MIT-licensed and behaves identically on a laptop and on a rented GPU box — the only thing that changes is how long you wait for each token. That waiting is the whole reason this page is about hardware.

See what a GPU costs → Browse all GPU plans
ollama — install & serve
$ curl -fsSL https://ollama.com/install.sh | sh
✓ daemon listening on :11434
$ ollama run gpt-oss:20b
✓ pulled, chatting in the terminal
$ curl localhost:11434/v1/chat/completions
✓ same shape your code already speaks

What Ollama actually does

MIT, no strings
The runner itself is MIT-licensed. What it runs is a separate question — each model carries its own licence, and a permissive runner does not make a restricted model permissive.
One command to a running model
An install line and ollama run. No Python environment to reconcile, no CUDA versions to argue with, no config file before the first token.
An OpenAI-shaped API on :11434
The daemon serves /v1/chat/completions, /v1/completions, /v1/embeddings, /v1/models and /v1/responses. Existing clients point at a different base URL and otherwise carry on. The API key is required by the clients and ignored by the server.
Nothing leaves the box
Weights sit on your disk, prompts stay in your process. There is no upstream to read a retention policy about, which for most people is the entire point of self-hosting.
Quantized by default
Models arrive already compressed, which is why a 20-billion-parameter model is a 14 GB download rather than a 42 GB one. That compression is what decides whether it fits the card you rented.
It runs without a GPU too
Ollama will happily use a plain CPU. It will also be slow enough that nobody uses it twice for anything interactive — which is the honest argument for renting a card rather than a bigger VPS.

How much GPU memory each size needs

This arithmetic is ours, not a vendor requirement. Ollama publishes download sizes but no per-model GPU minimum, so we do the sum in the open: weights take roughly 2 bytes per parameter at FP16, 1 byte at FP8 and 0.5 byte at 4-bit, plus +25% for the KV cache and the serving runtime. It is a floor. A long prompt, several people at once or an unquantized build all push it up. Ollama ships models quantized, so the 4-bit row is the one that describes a normal install.
Model size Weights at 4-bit With headroom Card that holds it What it is good for
1–4B ≈ 0.5–2 GB ≈ 1–2.5 GB Anything with a GPU Classification, extraction, drafting
7–8B ≈ 4 GB ≈ 5 GB A 6–8 GB card The usual general-purpose workhorse
13–14B ≈ 7 GB ≈ 9 GB A 12 GB card Noticeably better reasoning
30–32B ≈ 16 GB ≈ 20 GB A 24 GB card Where self-hosting starts to feel worth it
70B ≈ 35 GB ≈ 44 GB A 48 GB card, or two 24 GB The top of what one rented box does comfortably

Two published figures to calibrate against, both Ollama’s own rather than ours: gpt-oss:20b is a 14 GB download and gpt-oss:120b is 65 GB, each with a 128K context. A download size is not a memory requirement — the context and the runtime sit on top of it — but it is a real number from the people who ship the model, and our arithmetic above lands in the same neighbourhood. For the models themselves, see GPT-OSS and DeepSeek.

Cheapest GPU at each tier

All GPU plans →

The cheapest live GPU plan in the DohoHub catalogue at each memory tier, out of 142 plans whose provider publishes a GPU memory figure. Prices come from our own feed, normalised so they are comparable and shown in the currency picked in the header. Where one machine is the cheapest way to clear several tiers, it appears once and the note says which tiers it covers.

Provider Plan Specs Price Visit
Servarica
Bee GPU VPS
cheapest at 6–24 GB+ · 24 GB on board
 6 Cores24 GB vRAM30 GB RAM400 GB NVME100 TB traffic10 Gbps
$79.00/mo
 Visit
NovoServe
Supermicro X11 10SFF (GPU)
cheapest at 48 GB+ · 64 GB on board
 64 GB vRAM25 TB traffic
$355.99/mo
renews at $402.52
 Visit

We record the memory a plan advertises, not which card supplies it, and that matters more here than the number suggests: two older cards adding up to 24 GB will load the same model as one modern 24 GB card and generate tokens at a very different speed. None of these providers sells “Ollama hosting”. They sell GPU servers, and the install line above is the whole setup.

Pick a card that fits the model you want

Filter the catalogue by GPU memory, cores, storage and location — live prices, tracked every six hours.

Find a GPU server →
↑