A local model runs on your own GPU. No per-word cost, nothing leaves your machine, and you download it once (4 GB to 33 GB depending on the size). A cloud model runs on Ollama's servers. No download and no GPU needed, but you pay per token, and the conversation is sent to Ollama.
Both live in the same tray app: right-click the tray icon → Change model → Local or Cloud. Switching to a cloud model finishes in seconds because the image is already on disk. Switching to a local model downloads it first.
Input is everything you send to the model: your message, the system prompt, any file you paste in, and the earlier turns of the conversation that get sent back with each new message. It is billed per million tokens.
A token is roughly three quarters of a word — a thousand tokens is about 750 words.
Output is what the model writes back. It is usually the more expensive of the two, because generating text is the slow part — every word is computed one piece at a time.
One exchange costs (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price). As a rule of thumb: a short question with a long answer is dominated by the output price, and a long document you paste in is dominated by the input price.
Ollama has a cheaper rate for text it has already seen recently. It is real, and it is applied to your bill automatically, but it depends on how the provider happened to cache that request — two people asking the same thing can be charged differently. Showing it next to the two honest numbers only raised the question "which one am I actually paying?", so it is not listed. Your bill at ollama.com/settings is the exact number.
Loading the current list…
| Model | Input / 1M | Output / 1M |
|---|
Cheapest output first. Prices are in US dollars and come from the same list the app reads, so this table and the tray never disagree.
Yes, and not always with notice. The table above is fetched live from /api/cloud-models.json every time you load this page, so it is never a transcript of an old decision. What you were charged for a past conversation does not change retroactively.
The model's row disappears from the list. Not greyed out, not marked "unavailable" — gone, because a row you can click that cannot work is worse than no row. Your running container keeps working until you switch, and the tray tells you when a listed model has been retired so you can pick a replacement.
Nothing is substituted behind your back. We do not swap in a cheaper or older model under the same name; if a slot empties, it stays empty until a real model fills it.
Only for cloud models. Get one at ollama.com/settings/keys and paste it into the tray's key card the first time you pick a cloud model. It is checked against ollama.com before it is saved, so a typo fails on the spot instead of after a switch.
The key is stored in a file on your PC with owner-only permissions. It is mounted read-only into the container as a file — never passed as an environment variable — so docker inspect and a support log cannot leak it.
Your prompts, the conversation context, and anything you attach to a message. Not your other files, not your chat history that is not in the current context, and not your settings. Local models send nothing anywhere — that is the reason to keep one installed even if you mostly use cloud.
As often as you like. Local files and chat history survive the switch; only the model changes. The old image is kept so switching back is fast, and there is a Clean up models entry in the tray when you want that disk back.
Check the tray's status panel first — it shows which model is running, whether the container is healthy, and the last error in plain language. If that does not explain it, mail us at support@psyllama.com with the model name and what the status panel says.