July 20, 2026
Hot mic, cold trust
Model 4: Our best model yet for local dictation and cleanup on Mac
Your Mac gets smarter offline, but commenters are stuck on the pricing fine print
TLDR: Epilude says its new Mac dictation model is much better at cleaning up speech locally, so your voice never leaves your computer. But the loudest reaction wasn’t about the tech at all—it was confusion over whether “pay once” actually includes future model upgrades, and commenters clearly want answers.
A new dictation upgrade just dropped, and on paper it sounds like a privacy lover’s dream: your Mac can turn messy speech into polished writing without sending your voice to the internet. The company says Model 4 is its biggest leap yet, with fewer meaning-changing mistakes, better handling of corrections like “no wait, Thursday,” and the same download size and speed as before. In plain English, it’s supposed to be better at not making you sound like a robot—or accidentally rescheduling your life.
But in the comment section? The real drama isn’t just the software, it’s the money talk. One early reaction zeroed in on the website’s pricing language and instantly shifted the mood from “cool update” to “wait… what exactly am I buying?” That sparked the classic tech-community side eye: if there’s a subscription and a “pay once” option that only includes “one year of app updates,” people want receipts. Does that mean app updates only? Model updates too? Both? Neither? The vibe is less “wow, offline AI” and more “please explain this like I’m five before I click buy.”
There’s also a darkly funny undercurrent here: the product promises zero trust-breaking errors, while commenters are basically saying the same standard should apply to the pricing page. The unspoken meme of the moment is obvious: your dictation may be deterministic, but your billing better be too.
Key Points
- •Epilude released Model 4, a new generation of the on-device cleanup model used in Local Mode for Mac dictation.
- •The update keeps the same 1.5 GB download size, the same speed, offline operation, and unchanged speech model, while replacing the cleanup model.
- •Model 4 is Epilude’s fine-tuned model built on the Qwen family and is designed specifically for dictation cleanup.
- •Epilude evaluates models using an internal 90-scenario benchmark and focuses on preventing critical errors that change meaning or leak instructions.
- •On Epilude’s internal benchmark, Model 4 passed 78 scenarios versus 69 for Model 3 and reduced critical meaning errors from 9 to 4, with zero instruction leaks in a 220-input stress test.