Cotypist - Local AI Autocomplete in your voice, anywhere on your Mac

@nuseir_yassin1 Cotypist lets you provide custom instructions to generally tune it to your writing style as a starting point. But it indeed also learns from your writing over time, and that's where the real magic happens, to the point where it feels like a best friend completing your sentences — or as if it were reading your mind 😉

@syed_shayanur_rahman Thank you! Yes, I was very happy when they reached out about a relaunch.

@roopreddy Indeed! On the other hand, Cotypist being almost invisible does not stop user from relying on it; I've even heard from users who thought their Mac was broken when Cotypist was not active.

@lennarto Thank you for the kind words, I really appreciate it!

@krekeltronics Indeed! In my experience, Cotypist's suggestions at this point are so often relevant that it's better to err on the side of showing them. But there's still the challenge of making sure that you never have to wait for completions to appear, which is why I’ve put a lot of effort into optimizing Cotypist’s performance.

@shubham4real Hi Shubham, you've put your finger on exactly why I built it this way. Sending every keystroke to a server means your drafts and half-formed thoughts live in someone else's logs, and running inference locally is what avoids that completely. On your practical question, which is the right one to ask:

Model size. The default is Google's Gemma 4 E2B, running entirely on your Mac. It's a few gigabytes on disk, but only about 1 GB of it needs to be in active memory while it's generating. Even though it's a multi-billion-parameter model, Cotypist doesn't keep all of its weights resident at once: part of them are streamed in from disk only as they're needed, so they never have to sit in GPU memory. That's a big reason it punches above its weight, you get close to the quality of a much larger model for the footprint of a small one.

The quality vs. resources tradeoff. Cotypist matches the model to your hardware rather than running one heavy model everywhere. Lighter and older Macs get the roughly 1 GB E2B by default; stronger Macs (the M-series Max and Ultra chips) can step up to the larger E4B, around 2.5 GB in memory, for better suggestions. The heavier models are only offered on machines that can actually run them well, so you won't accidentally bog down an older Mac, and it winds down and frees that memory when you're not actively typing.

Speed. This is the part I've spent the most time on, because autocomplete only earns its place if it clears two bars at once: good enough that you usually want to accept it, and fast enough that you're never waiting on it. Suggestions appear in real time, usually with no noticeable delay, and they keep updating as you write. People run it comfortably even on a base MacBook Air. The honest test for your particular machine is the free trial, so you can feel the speed on your own hardware before deciding anything.

@neilio Thank you for sharing your experience! The goal has been for the free plan to still be genuinely useful and sufficient for many users, so it’s great to have that confirmed. Enjoy!

@dylanb Thank you for the endorsement! I've indeed put a lot of effort into making sure that completions update quickly as you type, so even if the initial suggestion isn’t quite right, typing just one or two more letters will often give you exactly the word you were looking for, so that Cotypist still saves you the effort of typing the second half of the word.

@phillip_b Glad to hear that! I am with you on how frequently I use it. Cotypist has already completed 3250 words for me today, which, at about 7.5 hours of computer usage, translates into roughly one word every 8 seconds 🤯

@david_weinberger4 Thank you for sharing your experience! I’m glad to hear that Cotypist is able to make suggestions that feel like they come from you.