Choosing a model
A model on this computer, your own key, or Plus + Credits; which Gemma size fits 8 GB and 16 GB; and how to change it in Settings → Model.
Updated 4 Oct 2026
Pivra Desk is not tied to one model. You choose where the model runs during setup, and you can change it any time in Settings → Model.
The two main choices
| Free and private · Runs on this computer | Faster · With a Claude or OpenAI API key | |
|---|---|---|
| Account | None. Nothing to pay. | An API key from Anthropic or OpenAI, paid as you go. A Claude Pro or ChatGPT Plus subscription does not include one. |
| Where your text goes | Nowhere. What you write never leaves this computer. | From your computer to that provider, under your own account and their terms. Not through Pivra. |
| Good at | Reading, summaries, drafting, short tasks with tools. | Long, multi-step work. |
| Speed | Depends on your computer. On an M2 Pro with 16 GB, Gemma 4 E2B writes about 78 tokens a second (measured 4 October 2026). | Fast; depends on the provider. |
Getting a key: in the provider's console, add some credit under Billing (five dollars is plenty to start), create a key under API keys, copy it, and paste it into Pivra Desk. Where do I get a … key? under each provider in Settings → Model has the exact steps.
Which model on this computer?
Offline mode runs Google's Gemma 4 or a small Qwen through llama.cpp on your own machine. Pivra Desk downloads the file once from Hugging Face and checks it against a pinned checksum before using it.
| Model | Download | Pivra Desk's own note |
|---|---|---|
| Gemma 4 E2B | 3.3 GB | Smaller Gemma for 8 GB computers that are already busy. Faster than E4B, less careful. |
| Gemma 4 E4B | 5.2 GB | The everyday choice for 8 to 16 GB computers: reading, summaries, drafting, and short tasks with tools. |
| Gemma 4 12B | 7 GB | Needs a lot of memory: more careful answers for computers with 24 GB or more. On 16 GB it pushes other apps out and can slow to a crawl. |
| Qwen2.5 0.5B | 0.5 GB | Tiny and fast, for computers with little memory. Short, simple tasks only; it often gets tool use wrong. Automatic memory stays off while it is in use, because it is too small to tell lasting facts from its own replies. |
Before it loads a model, the app checks the memory free right now and picks a context size that fits. If memory is tight it uses a shorter context; if it is short, the model still loads but crawls, and the app tells you to close other apps or choose a smaller model in Settings → Model. You can keep more than one model downloaded and switch between them; You can download it again later if you remove one.
On an 8 GB computer
Start with Gemma 4 E2B. Close the browser tabs and apps you are not using before a long task. If it is still slow, Qwen2.5 0.5B answers simple questions quickly but is not reliable with tools. We are measuring 8 GB Macs and Windows laptops and will publish the numbers.
Advanced: other model services
Under Advanced: other model services you can use OpenRouter, Cloudflare AI Gateway, Vercel AI Gateway, or any OpenAI-compatible endpoint, including a local server on your own network: Ollama (http://127.0.0.1:11434/v1) or LM Studio (http://127.0.0.1:1234/v1). Pick the Model from the list, or choose Other… and type its name.
Plus + Credits
The Plus + Credits plan (US$15 or A$23 a month) includes monthly cloud model credits so you can use a cloud model without a key of your own. The in-app credits model is still being built; until it ships, Plus + Credits gives you everything in Plus, and this page will say when the credits are usable. If you need a cloud model today, your own key is the way. See Plans.
Which should I pick?
- You want nothing to leave your computer: Runs on this computer, Gemma 4 E4B on 16 GB, E2B on 8 GB.
- You hand the agent long, multi-step jobs: your own Claude or OpenAI key.
- You already run Ollama or LM Studio: point Pivra Desk at it under Advanced.
Whatever you choose, approvals work the same way: the model proposes, and anything that sends, pays or changes something waits for you.
Was this helpful?
If you have questions or suggestions, email us at support@pivra.ai .