Check the model route, not just the server address
A frontend pointing at localhost:11434 tells you where its first request goes. It does not settle where the answer is generated. These three routes have different consequences.
On a narrow screen, scroll the table horizontally. Keyboard users can focus the table area and use the arrow keys.
| Route you select | Where inference runs | What to check |
|---|---|---|
| Downloaded local model Through your local server | Your machine | The exact installed model and hardware fit. No Ollama cloud-token bill; hardware, storage and electricity still cost money. |
Cloud model through localhostgpt-oss:120b-cloud | Ollama’s cloud | The local server signs in and forwards the request. Cloud usage terms and credits apply despite the local endpoint. |
Direct cloud APIhttps://ollama.com/api | Ollama’s cloud | An API key is required. No local installation is needed. Use the model name returned by the hosted model list. |
Sources: authentication and forwarding examples, cloud setup and model naming, local and cloud pricing.
A missing -cloud suffix is not a universal local-execution test. Hosted API model names differ from app and CLI names. Read the selected provider, endpoint and model together. Naming rules.
Cloud privacy promises are different from keeping text on your machine
Ollama’s March 2026 policy says it does not collect locally processed prompts and responses. For cloud models, it says content is processed transiently, without training on it; limited usage metadata and account information are still collected. These are vendor commitments, not findings from our own network inspection. Current privacy policy.
Disable cloud deliberately if local execution is the requirement
The default server binds to 127.0.0.1:11434, and its local API has no authentication requirement. Keep that boundary in mind when connecting other software; a local endpoint is not an account-protected service. Default binding; local authentication rules.
Ollama documents two cloud-off options: merge "disable_ollama_cloud": true into ~/.ollama/server.json, or OLLAMA_NO_CLOUD=1 in the server’s environment. Restart the running Ollama app or service and look for Ollama cloud disabled: true in its logs. This disables Ollama cloud models and web search. Local-only configuration.
That switch does not configure a separate frontend, its other providers, extensions or MCP tools. I would check those connections separately before placing a private manuscript in the workflow.
Suggested check · Not performedTry a fresh writing request with the network disconnected
- Download a compatible local text model first. Record its exact name and version, and read its license.
- Choose one documented cloud-off option, restart the actual server process and inspect the confirmation log. If editing an existing JSON configuration, preserve its other settings.
- In the frontend, select that local model and endpoint. Disable other cloud routes, automatic provider fallbacks and optional web or tool connections for this controlled check.
- Disconnect the network briefly, start a fresh conversation and supply a new benign fictional detail, such as “Mara’s blue notebook is named Cedar-47.” Ask for a short scene that uses it.
- Record the result and inspect
ollama pswhile the model is loaded. Reconnect when finished.
A fresh response using the new detail would support offline operation for that request and configuration. It would not prove that every feature stays offline when reconnected, that no other app transmits data, or that stored transcripts are encrypted. A failure could also come from a missing model or a frontend dependency.
Inspect allocated context before blaming character memory
Ollama’s dedicated context guide now describes VRAM-based defaults; its FAQ still states a single general default. Check the running allocation rather than assuming a documentation default. After loading a local model, use ollama ps to inspect CONTEXT and PROCESSOR. Current allocation guide; FAQ wording.
Increasing context uses more memory. A model’s advertised maximum is not the same as the allocation on your machine, and CPU/GPU placement matters when assessing a slow session. We have not benchmarked either. Context and offloading.
For a continuing fictional story, keep a short canon note in the frontend and inspect what it sends with each turn. Saving a transcript does not ensure the model receives all of it. A correct answer immediately after a reminder tests recent context, not durable persona memory.
I would fix an omitted or contradictory canon note before increasing context. If you want character cards and conversation controls around a model connection, compare the SillyTavern review; those controls introduce their own storage and provider settings.
Local use has no cloud-token bill; hosted use now spends credits
For new individual subscriptions, the current pricing page lists Pro at $20/month with $60 of monthly usage credits, and Max at $100/month with $300. Free accounts can buy usage credits without subscribing. Included credits reset rather than roll over. Check your own account before changing a plan. Current prices and credit rules.
Cloud rates distinguish input, cached input where offered, and output tokens. Price a complete writing session, including rewrites. If a frontend resends a long conversation on each request, that history matters to cost; a short new message is not necessarily a short billed input. Do not assume every repeated input earns the cached rate. Per-model token rates.
Check the plan transition before switching billing cycles
The August 31 pricing announcement says existing subscribers can keep their legacy plan through continued automatic renewal. A billing-cycle change, tier change or cancellation followed by resubscription moves them to the new pricing. Do not treat an older plan’s allowance as today’s offer. Legacy-plan transition rules.
Purchased usage credits expire after one year, a different rule from monthly included credits. May 2026 payment terms.