Loading video...
Video Failed to Load
1/ Today was the first time offline computer use felt real to us. We ran Muse Glimmer 30B from AI at Meta locally and watched it operate Notes and Reminders on macOS through Cua Driver We didn’t expect a 30B local model to get this far 🧵
30,795 views • 1 month ago •via X (Twitter)
28 Comments
Great stuff, thanks for sharing this feedback!
2/ Latency on consumer hardware like Apple Silicon still needs work, but we can now see where this is going. If speed isn’t the priority, Muse Glimmer is the first model i’d actually recommend for a fire-and-forget local agent workflow. It just keeps going until the job is done
3/ And here it is using Reminders:
4/ You can start by using @UnslothAI’s UD-Q4_K_XL quant with Claude Code as the harness. The biggest constraint right now is context. Tool schemas, screenshots and accessibility trees add up quickly for a local model. Docs:
5/ We kept expecting it to fall apart once it had to mix accessibility actions with pixel-grounded clicks. It did get confused in a few places, and the tool schemas mattered a lot. Once we tightened the interface, it completed and verified both tasks
6/ As a result, we’re going to make @trycua's Cua Driver friendlier to local models by cutting the context cost of its tools and computer state. The goal is to let smaller models spend more of their context on the task and keep running longer
7/ That’s what we appreciated about this release. Huge kudos to @alexandr_wang @finkd and the Muse Glimmer team at @AIatMeta for releasing a model people can run and learn from locally
You can now try offline computer use with cua-driver + Muse Glimmer, through our friends at @ollama 🦙 How-to:
@AIatMeta @jicapal @robotson lil guy bout to be so dexterous
@AIatMeta 本地模型在 macOS 上操作应用，离线代理体验到了临界点
@AIatMeta Did you guys try Qwen 3.7 27b?
@AIatMeta pretty good!!
@AIatMeta Thanks for sharing!
@AIatMeta Thats actually insane, really want to test it out. Keep going @francedot you are amazing!
@its6i77 @AIatMeta thank uuu 🫶
@AIatMeta 👀
Is there a list of model to cua performance benchmark somewhere? Side note: I just hooked up CUA into all my personal machines and I wish there was a macbar app so I could see when it was using my machine in the background. Is the PIP mode recommended for this? I really like how codex shows what is being used in the background.
@AIatMeta we're working on showcasing more of these in our docs. PiP it is experimental and we haven't dedicate much time since release, but would be good to find a better ux for this - @injaneity want to cook smth?
@AIatMeta The interesting part isn't just that a 30B model can run locally. It's that it can now operate a computer well enough to perform real tasks. Local models controlling real software changes the economics of computer-use agents dramatically. This is starting to feel very real.
@AIatMeta Offline computer use becomes much more interesting when you add privacy + latency benefits. For daily work, I’d still want action logs, permission boundaries, and rollback checkpoints before trusting it with real tasks.
@AIatMeta 30B能跑通demo和能日常用之间隔着一堵墙：上下文短、单步错误率固定，多步操作一长错误就指数累积。Cua这类本地驱动还得过macOS的TCC权限弹窗和辅助功能授权——这两层才是离线agent真正卡脖子的地方，不是模型智商。
@AIatMeta Open weights doing real work. You control the model, data stays local, and no classifier silently reroutes you to a weaker version.
@AIatMeta Huge for democratization. Travel, humanitarian work, and remote communities all benefit when useful AI doesn’t depend on a stable connection.
@AIatMeta With this kind of good local model, everyone is becoming a developer, and everyone has a database server by now, right? This is the perfect app for managing and controlling your SQL databases directly from Android ( or iOS (
@AIatMeta Been waiting for this. Local models going from 'writes text you paste somewhere' to actually operating the app. Whole different workflow.
@Dan_Jeffries1 @AIatMeta You did not have any successful results with prior models like gemma4-31b or qwen3.6-27B?
@AIatMeta Never ran Qwen3.6 27b then….
@AIatMeta That feeling when a local model actually works for real tasks is wild. We actually covered this capability shift here:
![Muse Glimmer one-shot 5 games! AI at Meta dropped Muse Glimmer. So we ran it against Muse Spark 1.2, on five one-file game demos, each one written to play itself: Tetris, a top-down pixel street race, a rooftop web-slinger, a blue hedgehog platformer, and Flappy Bird. the setup: • Muse Glimmer 30B — via aimlapi[.]com. cost: $0.02 • Muse Spark 1.2 — via aimlapi[.]com. cost: $0.11 what we saw: • every Glimmer scene runs and plays itself. it's a reasoning model, so it burns more tokens — but the output is clean. its Tetris even scores each placement to keep lines clearing, and the blocks burst into pixels on every clear. • Glimmer beat Spark at Flappy Bird — its bird flew through the pipes cleaner. • but it's not perfect: in the racing game its cars sometimes slip off the road, and its graphics are simpler and blockier than Spark's. glad to see Meta back in open source — and we're waiting for Spark 1.2 weights next 🙌](https://image.24vids.com/tw-2087109782570230146/amplify_video_thumb/2087109053080072192/img/ycepchkmjbhn7CyS.jpg)
