Models on your Mac

Canto's AI runs on your Mac. It picks a model by your Mac's chip and memory, says how long things will take, and never sends your notes to a model off your Mac unless you connect one. Everything here is in Settings › Models.

The recommendation#

This Mac, at the top of Settings › Models, shows your Mac, its memory in one bar, and the limit macOS lets models use drawn on that bar.

Under In use, the chat model's row is the recommendation for your Mac, with four figures measured for it here:

• Edit starts: how long before the first change of an edit lands.
• Web answer: about how long an answer with ten searches and two pages takes.
• Research: how long a research run takes, or "Use a server" where your Mac is better served by one.
• Memory: how much the model needs.

Other choices that suit your Mac are listed below it, each with what it's good at and what it costs here, such as "More careful writing · reads about 6× slower". The one that runs Thorough research says so. When your chat model isn't the one recommended, one row suggests the recommended model with its numbers; Not now hides it.

The rest of In use lists every job (quick jobs, chat, chat with web, editing a note, editing several notes, research, meetings, dictation, images and PDFs, search by meaning) with what runs it and how well: Runs well, Works, Slower, Very slow or Use a server.

What Canto recommends, by memory#

Canto first checks what fits (the graphics memory macOS allows and the memory free now), then that it's fast enough on your chip. In beta 3:

On bigger Macs, Qwen 3.8 27B is one click away for slower, more careful work, and runs Thorough research where it finishes in under 15 minutes.

On battery, in Low Power Mode or when the Mac is hot, long answers use a faster model and pause between steps, and say so. When memory is tight right now, a run steps down for that run only.

Choosing and loading a model#

Choosing a model, in Settings or in the composer's model menu, never loads it or starts a download. A model loads when you send a message, and the chat says so ("Loading Qwen 3.5 9B · 40%").

A chat starts with a small memory and grows as it needs to, so a short question doesn't reserve room for the longest possible chat. Canto unloads models it loaded after a few idle minutes and when you quit.

Background jobs, such as titles, tags and tidying suggestions, run only on models built into Canto (or Apple Intelligence), never on a server or a cloud model.

Memory: fits, not right now, won't fit#

Every model shows how much memory it needs, with a coloured dot:

• Green: it fits this Mac.
• Amber: it fits this Mac, but not right now. Quit a few apps, or wait for memory to free up. The model stays available, with a note.
• Red: it won't fit this Mac; it needs more than the graphics memory macOS allows.

Settings and the composer's model list always agree on this.

Downloads and builds#

On this Mac lists every built-in model in one list: the ones you have first, then the ones to get, then the ones that won't fit. Models come from Hugging Face, only when you press Get.

• A download shows how much, how fast and how long, and can be cancelled. Canto checks free disk space first.
• Each model has builds: Compact, Balanced or High precision. The builds of one model are one row; a popover shows which you have and which suits your Mac.
• Canto tells you when a newer upload of a model you downloaded is available.
• Delete a model from its row to get its disk space back.

Small Mac mode#

On a Mac with 12 GB of memory or less, Canto runs in Small Mac mode: shorter instructions and compact memory so Canto fits beside your other apps. There's no banner; Settings › Models says so in one line.

In Small Mac mode:

• quick jobs (titles, tags, short summaries) use Apple Intelligence where your Mac has it,
• chats use a lean setup, and the chat model unloads after 5 minutes idle,
• long research is better on a server or a cloud model, or answered from 10 web sources,
• research notes skip the figures pass for models on your Mac.

It's automatic. To change it, open Settings › Models › Advanced › Small Mac mode: Automatic (on for 12 GB or less), On or Off.

Apple on this Mac#

Apple on this Mac shows Apple Intelligence, Apple Vision and Apple speech, with their state, languages and help. It has two switches:

• Use Apple Intelligence for titles, tags and summaries: the quick helper, with nothing to download.
• Image descriptions: macOS describes each image in your notes, on macOS 27 with Apple Intelligence on. Before macOS 27, Canto reads the text and labels in images but not what a picture shows.

Text in images and scanned PDFs is read by macOS's Vision framework on your Mac, and audio and video are transcribed on your Mac. Nothing is uploaded.

Servers and cloud models#

Servers and Cloud on the same page connect models that don't run inside Canto: Ollama, LM Studio, oMLX, llama.cpp's server or any OpenAI-compatible server, and OpenRouter. A server on your Mac keeps your notes on your Mac; a server elsewhere or a cloud model receives what you send in those chats. Servers and cloud models covers them.

Advanced#

Under ▸ Advanced, Loading and context has:

Settings › Advanced › Performance has Low Memory Mode, which unloads models sooner and keeps smaller caches, for Macs with 8 or 16 GB.

If a model won't load or is slow#

See Troubleshooting › Models and memory.