Large models can exceed the GPU or system memory available on a local machine. Ollama cloud models offload inference to hosted hardware while the local client sends prompts and receives answers.
A running Ollama server, curl, and an Ollama account signed in on that server are required. Cloud inference consumes account usage credits and depends on the account's access to the selected model.
Prompts leave the local machine for hosted inference. Sensitive data should only be sent when that use is permitted; environments that require local-only inference need a local model instead.
Steps to run an Ollama cloud model:
- List the models available to the local Ollama server.
$ ollama listAn empty list means no model references have been added yet. Account sign-in belongs to the server that handles the request.
Related: How to sign in to Ollama - Pull the cloud reference for gpt-oss:120b-cloud.
$ ollama pull gpt-oss:120b-cloud ##### snipped ##### successThe cloud reference contains metadata pointing to the hosted model, rather than its weights. A local GPU and enough memory for the full model are not required.
Related: How to pull an Ollama model - Run the cloud model with a short arithmetic prompt.
$ ollama run --hidethinking --think=low gpt-oss:120b-cloud "What is 17 plus 25? Answer with only the number." 42
For this model, –think=low reduces reasoning effort and –hidethinking hides reasoning text from the terminal. An account-access or credit error means inference did not complete.
- Verify a reply from the same cloud model through the local API.
$ curl --fail-with-body --silent --show-error http://localhost:11434/api/chat -H 'Content-Type: application/json' -d '{ "model": "gpt-oss:120b-cloud", "messages": [ { "role": "user", "content": "What is 17 plus 25? Answer with only the number." } ], "stream": false, "think": "low", "options": { "num_predict": 64 } }' {"model":"gpt-oss:120b","created_at":"2026-11-04T02:05:01.831733866Z","message":{"role":"assistant","content":"42","thinking":"Just output 42."},"done":true,"done_reason":"stop","total_duration":863009103,"prompt_eval_count":81,"prompt_eval_cached_count":64,"eval_count":16}
stream: false returns one response, and num_predict bounds generated tokens. The assistant answer is in message.content; done: true with done_reason: stop indicates a completed answer. The returned model name can omit the local -cloud suffix. The local server uses its account sign-in for cloud access, so this request needs no hosted API key.
Related: How to authenticate Ollama API requests