FlowMail runs its language model on your own Mac. There is no API key to paste, no account to create, and no AI service on the other end of it. This page is what that actually means — which models, what they cost you in disk space, what the machine has to be, and what still goes over the network.
What “local” means here
The model file sits on your disk and the inference runs on your Mac’s own silicon, through Apple’s MLX. When a workflow summarises an email or drafts a reply, the text of that email is handed to a process on your machine. It does not go to OpenAI, to Anthropic, to us, or anywhere else.
The practical consequences matter more than the principle. There is no key to rotate, no per-token bill, no rate limit, and no outage. It also works on a plane.
The models you can choose
Four, picked in Settings. All are 4-bit quantised builds from the mlx-community collection on Hugging Face:
| Model | Download | |
|---|---|---|
| Qwen3 4B Instruct (2507) | ~2.4 GB | the default |
| Llama 3.2 3B Instruct | ~1.8 GB | |
| Qwen3 1.7B | ~1.0 GB | |
| Llama 3.2 1B Instruct | ~0.7 GB | the one for an 8 GB machine |
The list is short on purpose. It holds only the model families the shipped MLX runtime actually supports, so nothing in the picker fails to load after you have waited for a two-gigabyte download. Newer families are left out until the underlying package registers them — a picker full of models that might work is worse than a short list that does.
Apple Silicon only — and on an Intel Mac it is not there at all
MLX is Apple-Silicon-only, so every part of FlowMail that loads or runs a model is compiled behind an arm64 check. On an Intel Mac the AI steps are not disabled — they are not in the binary. Everything else works: the list, the Kanban board, the conversation graph, accounts, search, Shortcuts.
If you are on Apple Silicon, any M-series Mac will run the smaller models. The 4B default is happier with memory to spare.
Nothing is bundled, and you can switch it all off
FlowMail ships with no model inside it. You pick one in Settings and it downloads then — shipping a 2.4 GB file inside a mail client would be a strange thing to force on someone who only wanted to read their email.
There is a single switch in Settings that turns the whole AI side off. Download a model, delete it later, or never download one at all; FlowMail is a working mail client either way.
What the model is actually asked to do
It is not a chatbot bolted to your inbox. The model is one kind of step inside a workflow — an ordered list where each step can pass its result to the next as a named variable. The steps are:
- Prompt, Summarize and Classify — the three that run the local model
- Query emails — search your own mail by subject, sender or body
- Draft reply — see below
- Set Kanban status, apply label, remove label — move the mail around
- Web search and URL content — pull something in from outside
- Call MCP tool — reach a Model Context Protocol server you have configured
It writes drafts. It does not send them
A workflow can compose a reply. What it then does is save that reply into your Drafts, addressed to the original sender and threaded to the message it answers, and stop. There is no setting that makes it send. An automation that mails people on your behalf while you are not watching is a bad idea however good the model is, and it is the one mistake you cannot take back.
One honest limit in 26.1: that draft is written to the Drafts mailbox on this Mac. Pushing it up to your mail server, so it appears in Drafts on your phone too, is a follow-up rather than something this version does.
What does leave your Mac
“Local AI” is a claim about where the model runs, not a claim that the app never touches the network. Being precise about the difference:
- Downloading a model fetches it from Hugging Face, once, when you choose it.
- The Web search and URL content steps reach the internet by design — that is what they are for. Leave them out of a workflow and nothing in it goes out.
- A Call MCP tool step talks to whichever server you configured, with credentials you put in the Keychain.
- Your mail syncs with your own provider over IMAP or OAuth, as any mail client must.
What never happens is your email being sent to an AI service to be processed. FlowMail has no server of its own — there is no red8.io account and nothing is relayed through us, because there is no “through us” for it to pass. The FlowMail privacy policy has the detail.
What you need
- macOS 15.6 or later. FlowMail is a Mac app; there is no iPhone version.
- Apple Silicon, for the AI steps only.
- 0.7 to 2.4 GB of disk for a model, and only if you want one.
Running the local model is free. The Prompt, Summarize and Draft Reply steps all work without paying anything, as does your first workflow. FlowMail Pro adds the AI Classify step, unlimited workflows, the conversation graph and the thread tree.
More
- FlowMail — the three views, the board and the graph
- What is new in 26.1 — the release this page describes
- FlowMail FAQ — accounts, syncing, and what the AI will and will not touch
- How to choose the best email client for Mac — the seven questions worth asking, including where the AI runs and who pays per message