Ollama Cloud: what it costs, what it runs, and when to run Ollama yourself

What Ollama Cloud is

Ollama Cloud is the same Ollama app, command line and API, with models that run on Ollama's servers instead of your GPU. You sign in, pick a model whose name ends in -cloud or :cloud, and it answers as a local model would. Ollama launched it in preview on 19 September 2025. On 31 August 2026 it replaced GPU-time limits with per-token billing, because, in its own words, "GPU-time based billing was difficult to predict".

It serves models as "native weights, as released by the model provider", though Ollama says it may use NVFP4, a 4-bit format, on NVIDIA's Blackwell hardware. The list is Ollama's: 17 models today (18 counting both sizes of gpt-oss), and about half of them too large for any single computer you can buy.

Plans and credits

Every plan includes a monthly allowance of usage credits. Pro and Max give you 3 dollars of usage for every dollar you pay, and Pro billed yearly gives 3.6. Past the allowance you pay the same per-token prices from credits you buy; Ollama does not publish what a purchased credit costs, so the 3× is certain only inside the plan.

• Prices are before tax. Ollama's terms say "You are responsible for all applicable taxes".
• Included credits do not roll over. They refresh each month on your billing date. Credits you buy last a year: "Purchased usage credits expire one year after they are added to your account".
• No hard monthly cap. "when you use it up, you can keep going at the same per-token rate", drawing on purchased credits. Paid plans email you at 90% of the allowance.
• Older Pro, Max and Team subscriptions keep their five-hour and weekly limits while they auto-renew. Changing tier or billing cycle, or cancelling and subscribing again, moves you to per-token pricing; Ollama says support can switch you back.

What your credits buy

For a chat reply, a dollar of credits goes 56 times further on the cheapest model than on the dearest, so the model you pick matters more than the plan. The calculator prices three kinds of work at Ollama's published rates. They are worked examples rather than measurements: a chat reply (2,000 tokens in, 500 out), one step of a coding agent (30,000 tokens of context, 80% of it cached from the previous step, 800 out) and a long document summarised once (50,000 in, 1,000 out).

Ollama Cloud models and prices

US dollars per million tokens, as Ollama lists them on 8 October 2026. Cached input is input the model has already seen in a recent request, which agents and long chats reuse heavily.

* Nemotron 3 Super: input price as Ollama lists it; it is below the smaller Nano model's. The calculator ranks it at this price.

Which cloud models fit at home

About half of Ollama's cloud list is there because no machine you can buy holds it. Each row below gives the memory the model needs and the smallest machine in our data that holds it within 95% of its GPU memory, the rule the rest of the site uses. Models in our own data are computed at 4 bits (Q4_K_M) with 8k of context; for the two Nemotrons we use Ollama's 4-bit download size, which is the weights alone. For cloud-only models the figure is a floor: 4-bit weights take half a byte per parameter before any context, and every expert of a mixture-of-experts model has to sit in memory, not only the active ones.

5 fit a machine you can buy. 4 more fit on paper, since their figure is a floor and a model that only just fits leaves little room for context. 3 would fill the 384 GB the largest machine gives its GPU with nothing left over, and 5 need more than that. The largest machine is the Mac Studio M5 Ultra with 512 GB, a configuration that ships in late October 2026. A 4-bit copy at home is also not necessarily the weights Ollama serves.

Your own card or Ollama Cloud, per token

For the cloud models that also run on a single graphics card, the honest comparison is electricity against Ollama's output price. The break-even speed is the generation speed at which the card's electricity, at its full board power and $0.20 per kWh, costs the same as Ollama Cloud's price per token. Generate faster and the card is cheaper. It needs no speed estimate, only the board power and the two prices, which is why it leads; our estimate sits beside it.

Both assumptions favour the cloud: a card rarely draws its full board power while generating, and the cloud also charges for input tokens, which this leaves out. Electricity is also only part of the cost: the card itself is the larger one, and whether it pays for itself depends on how many hours you use it. The build vs rent calculator works that out for your numbers. Estimates marked ≤ are ceilings from memory bandwidth; real runtimes land below them, most of all for mixture-of-experts models that read few weights per token.

Privacy: what Ollama says, and what it means

• Logging and training: "Prompt or response data is never logged or trained on." The terms add: "We do not use your inputs or outputs to train AI models."
• Where it runs: "primarily in the United States", and it "may route to Europe and Singapore".
• Who runs it: "Ollama collaborates with NVIDIA Cloud Providers (NCPs) to host open models." Of those partners, Ollama says "we require no logging, no training, and zero data retention policies". It does not publish their names.
• The privacy policy is more careful than the pricing page. It says Ollama processes prompts "transiently", uses "technical measures designed to minimize retention of prompt and response content", and lists "model inference providers" among the third parties it shares data with.

Read together: Ollama makes a clear promise, but it is a policy you take on trust, not something you can verify, and your prompts pass through companies it does not name. Locally it is different by design: "We don't see your prompts or data when you run locally."

Running it yourself has its own risk. Researchers at SentinelLABS and Censys counted 175,108 Ollama servers reachable from the internet over 293 days of scans (a secondary source, not Ollama's). Ollama listens only on 127.0.0.1 by default, and its local API has no login, so a server opened to the network with OLLAMA_HOST=0.0.0.0 is open to anyone who finds it.

Limits and reliability

• Requests at once: 1 on Free, 3 on Pro, 10 on Max and Team. More queue up to a fixed limit and are then rejected. No other rate limits are published; the API answers 429 when one is hit and 502 when a cloud model cannot be reached.
• No status page and no uptime promise. We found no status page, and the terms say, in capitals, "we do not warrant that our services will be uninterrupted".
• Users have reported bad spells on Ollama's issue tracker: frequent 503 errors (April 2026), a "95% failure rate" (April 2026) and "model unavailable" errors (September 2026). 2 of these reports predate the August pricing change.
• Models are retired. Two of the four launch models of September 2025, Qwen3-Coder 480B and DeepSeek V3.1, are already gone; your settings show retirement dates for models you have used.
• Context: "Cloud models are set to their maximum context length by default." Locally, Ollama sets context by your GPU's memory, and its docs say agents and coding tools "should be set to at least 64000 tokens".

How to use it

From the app or the command line, sign in once and run a cloud model by name:

ollama signin
ollama run gpt-oss:120b-cloud

Coding tools start the same way: ollama launch claude, ollama launch codex and ollama launch opencode open Claude Code, Codex or OpenCode on a cloud or local model. From code, create an API key in your account settings and call ollama.com directly. The OpenAI-compatible endpoint is https://ollama.com/v1; Anthropic-style clients use /v1/messages.

curl https://ollama.com/api/chat \
  -H "Authorization: Bearer $OLLAMA_API_KEY" \
  -d '{"model": "gpt-oss:120b", "messages": [{"role": "user", "content": "Hello"}], "stream": false}'

curl "https://ollama.com/api/usage?range=24h" -H "Authorization: Bearer $OLLAMA_API_KEY"
curl https://ollama.com/api/balance -H "Authorization: Bearer $OLLAMA_API_KEY"

OLLAMA_NO_CLOUD=1 ollama serve

Names differ by route: the app and CLI take gpt-oss:120b-cloud, the API takes gpt-oss:120b. The current list is at ollama.com/api/tags. Ollama 0.40.1 (7 October 2026) was the latest stable release as we wrote this.

When to run Ollama yourself, and where

Run it yourself when the model fits your hardware, when prompts must stay with you, or when you want models Ollama Cloud does not carry: any open model on Hugging Face, your own fine-tunes, and adapters. To see what fits your card, open the can-I-run-it checker or a model's page, such as gpt-oss 120B or Gemma 4 31B.

If your machine is too small, the middle ground is a GPU you rent by the hour and run Ollama on yourself. A 96 GB card holds models up to gpt-oss 120B (71.4 GB) and Nemotron 3 Super (87 GB); the cloud GPU price guide lists the cheapest hour for each card. For DeepSeek V4, Kimi K3 and GLM-5.3, Ollama Cloud is still the answer.

Questions

Is Ollama Cloud free?

There is a free plan with a small monthly "starter" allowance for a set of starter models; Ollama publishes neither the amount nor the list. Every plan, the free one included, can buy credits, which unlocks every model without a subscription. Running models on your own hardware with Ollama is free and, in Ollama's words, "always unlimited".

How much does Ollama Cloud cost?

Pro is $20 a month ($200 a year) with $60 of usage credits; Max is $100 with $300. Usage is billed per token at each model's price, from $0.24 to $15.00 per million output tokens. Prices are before tax.

How many tokens does the $20 Pro plan give?

Pro's $60 of credits buys about 250 million output tokens on Nemotron 3 Nano 30B ($0.24 per million) or 4 million on Kimi K3 ($15.00). In chat replies of 2,000 tokens in and 500 out, that is about 250,000 on Nemotron 3 Nano 30B and 4,400 on Kimi K3.

Which models are on Ollama Cloud?

17 models (18 counting both sizes of gpt-oss), among them DeepSeek V4 Pro and V4.1 Flash, Kimi K3, GLM-5.3, Mistral Large 4, MiniMax M3, Nemotron 3, Gemma 4 and gpt-oss. The list changes: models are added and retired, and the current one is at ollama.com/api/tags.

Is Ollama Cloud private?

Ollama says prompt and response data "is never logged or trained on", and requires the same of the NVIDIA cloud providers that host the models. That is a policy, not something you can check: your prompts are processed on servers you do not control, mainly in the United States, sometimes in Europe or Singapore. For prompts that must not leave your machine, run the model locally.

Does Ollama Cloud train on my data?

No, by its own terms: "We do not use your inputs or outputs to train AI models."

Where are Ollama Cloud's servers?

Ollama says requests run "primarily in the United States" and may route to Europe and Singapore. We found no way to choose a region.

Can I run Ollama Cloud models locally?

4 of them have local downloads on Ollama: Nemotron 3 Super, Nemotron 3 Nano, Gemma 4, gpt-oss. Of the cloud-only rest, 5 need more than 384 GB for 4-bit weights alone, and the others fit only the very largest machines, on paper.

How do I get an Ollama Cloud API key?

Create one in your account settings on ollama.com and send it as a Bearer token. Keys do not expire until you revoke them. Through the app or CLI you sign in instead, with ollama signin.

Is Ollama Cloud OpenAI-compatible?

Yes, at https://ollama.com/v1 with your API key, and Anthropic-compatible at /v1/messages. Ollama says the cloud API supports only part of OpenAI's: no stateful Responses, and no built-in web search through /v1/responses.

Does Ollama Cloud have usage limits?

For new and changed subscriptions, the five-hour and weekly limits went with the move to per-token billing on 31 August 2026; older subscriptions keep them until you change tier or billing cycle. What remains is how many requests run at once: 1 on Free, 3 on Pro, 10 on Max. When the monthly credits run out, usage continues from purchased credits.

Can I use my own or fine-tuned model on Ollama Cloud?

Not as far as we can tell. Ollama's cloud offers the models on its list, and we found nothing on uploading weights or adapters in its pricing page, cloud docs, import docs or FAQ. Your own models run on Ollama locally.

Is Ollama Cloud worth it?

For models no home machine can hold, it is the simplest way to use them from the tools you already have. For models that fit your GPU, local is unlimited, and for gpt-oss the card's electricity costs less per token than the cloud's price above a modest speed (see the table above).