Four homes for the brain
Every AI feature (chat, email insights, action suggestions, builds) runs on one model: the default entry in Settings → AI Assistant. It can live in four places, and the choice is always yours:
- This Mac (default). An open-weight model like Gemma or Qwen via the built-in llama.cpp engine. Free, offline, nothing leaves the machine. Offered on Macs with 8 GB of memory or more, sized to the machine: on an 8-16 GB Mac, nenva's own small Qwen model, tuned for email insights and the assistant's tools (light and fast); larger open-weight models from 32 GB up. The smaller the model, the simpler the tasks it handles well; nenva Cloud is the stronger option for a small Mac, and setup suggests it below 8 GB.
- A server you own. Any OpenAI-compatible endpoint you host:
llama-server, vLLM, LM Studio on a homelab box. Good when your Mac is light on RAM but another machine you own isn't. - A provider you trust, with your own key. The official OpenAI or Anthropic API, added as a model entry with a key from your own account. Frontier capability that no local model matches yet, at the cost of sending what runs on that model to the provider.
- nenva Cloud. An open-weight model hosted by nenva, for Macs that can't run a capable model: no account, no key, a free monthly allowance of AI requests. What runs on it goes to nenva's server (api.anjadhe.com), which forwards it to a zero-data-retention inference provider, without your identity. Nothing is stored and nothing is used to train models; the service keeps usage counts, never what you asked, and its source code is public so that can be checked.
Every off-Mac option is an explicit choice: nothing ever falls back to a provider you didn't add, and your traffic goes only where you pointed it.
When the model runs off your Mac, Settings › AI Assistant › Cloud Privacy decides which kinds of data background work may send (journal and wellness stay home by default), and LLM Logs lists every request that left. The details are in Cloud privacy.
Adding a cloud model
- Settings → AI Assistant → + Add model. The page lists the homes on the left: On this Mac first when your Mac can run one, nenva Cloud first on a Mac too small for any local model.
- Pick nenva Cloud, OpenAI API, or Anthropic API.
- nenva Cloud needs no key: pick a model when more than one is offered (each shows a short description of what it's good at) and click Add; you can add several and switch between them. For OpenAI/Anthropic, paste an API key from your account (platform.openai.com/api-keys or console.anthropic.com/settings/keys), click List models to fetch the live list your key can use, and pick one (or type a model id).
- Test if you like, then Add model.
Make it the default via the radio on its card, and it becomes the brain for everything. You can keep local and cloud models side by side and switch from the model chip in the chat box: a local model for everyday private work, a frontier model when you need the horsepower. Keys are stored encrypted on your Mac, per model, and never sync; each Mac needs its own copy. The nenva Cloud card's Manage panel shows how much of the monthly allowance is used.
Pointing one part of the app at a different model
By default everything runs on your default model, the AI Assistant included. Once you have a second model, Which model answers what appears at the bottom of Settings → AI Assistant → Models and lets you send one part somewhere else: Email and texts, News, Documents, Portfolio, Routines and tasks, or Small jobs (memory, filing a task, suggested prompts). A small fast model can read your mail while a larger one answers your chat; your own API key can handle documents while everything else stays on your Mac.
Chat itself is not in that list: it runs on whichever model is marked Default in the list above, which is also what the model chip in the chat box switches. Every row left on Follows the default model moves with it, so the list only holds what you deliberately sent elsewhere.
Two things worth knowing:
- Your Mac runs one local model at a time. A second local model is offered only in place of the first; switching between two of them would reload gigabytes of weights on every request.
- Work sent to a model off your Mac runs beside your chat instead of waiting in line behind it, because it is not competing for your Mac's memory. That is the one way to make nenva do two things at once.
What each part may send off your Mac is still answered by Cloud privacy, and it is asked about the model that part runs on, so a part you keep on your Mac stays home even when your chat model does not.
When a cloud model makes sense
On an 8 to 16 GB Mac the local models on offer are small and fast, and they handle everyday tasks; for harder work (long documents, multi-step tasks), nenva Cloud is the no-setup upgrade, with a free monthly allowance. When it runs out, requests pause until the 1st. If you already trust OpenAI or Anthropic with the kind of questions you'd ask a chatbot, bringing your own key gives the assistant a much stronger brain for exactly those conversations, while everything else about nenva (your data on disk, sync, accounts) stays local and unchanged.
With your own key, be clear-eyed:
- What you run on that model goes to that provider under your account and their data terms. A chat that includes your briefing, or an email insights pass, sends that content to OpenAI or Anthropic.
- Usage is billed by the provider to you. nenva adds nothing on top; you pay OpenAI/Anthropic directly at their API rates.