Update, September 2026: small Macs run local AI again. Since late August, setup offers nenva's own small Qwen model, tuned for the assistant's tools, on any Mac with 8 GB of memory or more, with larger open-weight models from 32 GB. nenva Cloud remains the stronger brain for a small Mac, and everything below about how it works still holds. The next few paragraphs describe where things stood in August.
The most private AI assistant is one that runs entirely on your own Mac. That is still nenva's default, and on a Mac with 32 GB of memory or more it works well.
But we recently had to get honest about smaller Macs. The models that fit in 8 or 16 GB of memory never worked well enough to stand behind: they miss tool calls, invent answers, and quietly teach you that the assistant is not worth talking to. So nenva stopped offering them. That left an obvious question. What about everyone on a MacBook Air?
Until now the answer was "bring your own server or your own OpenAI or Anthropic key", which most people do not have and should not need.
nenva Cloud
nenva Cloud is a capable open-weight model that we host, for Macs that cannot run one locally. No account, no key, no card. During setup, if your Mac is short on memory, it is offered as one of your options. Say yes and the assistant works. While it is in preview, 1,000 AI requests a month are included, free.
It is the same deal we made with web search, applied to the harder problem: if a feature has to touch a server, say exactly what leaves your machine, keep the sensitive part out of our own hands by design, and publish the code so you can check.
What happens to what you send
With nenva Cloud selected, your chats, and the local data the assistant reads to answer you, go from your Mac to nenva Connect, the small server we run. It forwards them to an inference provider that runs the model, and brings the answer back. If you also turn on AI email features, the sender, subject and body of incoming messages make the same trip so the app can summarize them. That is the whole trade, and four things make it a fair one:
- We do not store what you send. Prompts and answers are never written to our logs or kept on our server. It counts your requests and tokens against the free allowance, and that is all it keeps.
- We do not know who you are. Your Mac uses the same random token as built-in search. No account, no email, no name. There is nothing to attach a conversation to.
- The model provider runs zero-retention inference. What the assistant sends is not stored and not used to train models, under the provider's own retention terms, and it arrives from our one server with all users mixed together.
- You do not have to take my word for it. The server's source code is public, and the app's privacy policy names this path in plain language.
Never a default, never a fallback
nenva never sends your data to a cloud model you did not choose. If you never add nenva Cloud, it does not exist for you: nothing falls back to it when the local model is busy, and nothing tries it quietly. Whatever model you select is the brain for every AI feature, chat and email analysis alike, so the choice you make is honest and whole.
Where the intelligence runs is yours to decide
- On your Mac, the default, and the most private. An open-weight model running locally; nothing leaves the machine. nenva's own small model on 8 to 16 GB, larger models from 32 GB.
- nenva Cloud, for Macs that cannot do that. Free monthly allowance, no account. Paid capacity may come later.
- Your own server, an OpenAI-compatible endpoint you host and control. Requests go straight from your Mac to it.
- Your own key, for an OpenAI or Anthropic frontier model. Straight from your Mac to that provider, governed by your agreement with them.
Private by default, and honest about the rest. Your data goes only where you point it.