8B is comfortable; 14B works if you close other heavy apps.
With the chanio starter installed:
chanio init --model qwen3:8b
Switching models keeps your notes, saved pastes, morning summaries and meter exactly as they are. The model is rented; the mind is yours.
Download sizes are Ollama’s default 4-bit tags as listed on ollama.com/library in September 2026. The memory rule of thumb (about 0.6 GB per billion parameters at 4-bit, plus 2–6 GB for context, plus 4–6 GB for macOS) comes from the community hardware guides; speed depends on the chip and on how much else is open.
The model is rented. The mind is yours.
Every model on this page will be replaced by a better one within a year. Your notes, saved pastes and morning summaries will not be. chanio keeps them as plain Markdown in ~/chanio so the next model reads the same files the last one did. One line sets it up.
Questions people ask
Why does memory matter more than the chip?
On Apple silicon the model lives in unified memory shared with macOS. If the model plus its context window do not fit, it swaps to disk and slows to a crawl. A faster chip makes a model that fits run faster; it does not make a model that does not fit run at all.
What does 3B, 8B, 27B mean?
Billions of parameters. Bigger models read more carefully and reason better, and need proportionally more memory and time. At 4-bit quantization, budget about 0.6 GB per billion parameters plus room for the context window.
Is a 3B model actually useful?
For answering from your own notes, summarizing a day, and drafting, yes. For hard reasoning, current events, or long generation, no; that is what keeping a question for tonight is for. The honest setup is a small AI that knows when to stop and keep the question.
Can I run a bigger model on an 8 GB Mac?
You can pull it; it will swap. A 7B model is borderline at 8 GB and a 14B is not practical. Stay with 3–4B models and keep the context window modest.
Which model does chanio use by default?
The starter looks at your Mac's memory: with 16 GB or more it installs gemma4:e4b (Google), which was right 48 times and wrong 4 on our 53-question test, against 43 and 6 for the smaller one; below 16 GB it installs llama3.2:3b (Meta), because it fits every modern Mac. Switch any time with chanio init --model; nothing else changes.
Who makes these models, and do they see my notes?
Llama is Meta's (US), Gemma is Google's (US), gpt-oss is OpenAI's (US), and Qwen is Alibaba's (China). Each card names its maker. They do not see your notes: a model here is a file of numbers that Ollama runs on your Mac, and chanio sends your notes and questions only to Ollama at localhost:11434, on the same machine. If a maker's country matters to you, every memory size above has a Meta or Google choice.