Beta 3 has dropped!
The current industry standard for AI agents is broken. The default architecture is to have a massive model sit there for 15 seconds burning tokens on an internal planning loop just to figure out what tool to use. And after all that waiting, it still hallucinates or picks the wrong thing too many times.
We threw that approach out. Beta 3 separates routing from the actual work.
A tiny 0.6B neural model runs directly in unified memory on your Mac, classifying user intent in about 20 milliseconds with typed accuracy. If you ask a coding question or just want an answer, it skips the planning theater entirely and streams in under half a second.
The speed boost is huge, but the real win is that it actually gets the routing right without doing dumb things like searching your drive for a file you literally just dragged into the window. Plus, you can now sideload any GGUF model you want directly from your SSD.
Still free. Still 27 MB. No subscriptions, and nothing leaves your Mac. https://t.co/szMMSAVvag
Just watch, we'll have regressed some basic feature when we implemented this, so now you can't even chat with an agent or something stupid. J/k, of course we tested it...
Just look at these release notes for 0.2.19. Everyone complains that AI is such a security risk. We've done pretty well with mitigating a lot of that risk:
Iβm sorry, itβs a fun product, but if you hand all your data to Muse and donβt expect it to be used for the most invasive things imaginable you are Charlie Brown with the football and you deserve whatever happens to you.
@WIRED You can protect yourself by downloading OOMU today because everything stays on your Macbook, there's no accounts, no telemetry, nada. And, it's free.