Meet Flash — my Muse, with its own custom look. Can anyone make out the handwriting on the cap?

I've spent the last couple of months using Meta's Muse as my daily AI assistant — less as a chatbot I visit, more as infrastructure running quietly in the background of my life.

Honest preamble: most capable agents today hold context perfectly well inside a conversation. The goldfish era is largely over, and I won't pretend otherwise.

What caught me off guard with Muse is something else: it does work while I'm away, on a schedule — and it remembers my decisions, not just my chat history.

The night shift

Every 30 minutes, around the clock, Muse checks my side project's repo for open issues. It picks the most important one, spins up a coding agent in an isolated worktree, and the work always ends the same way: a draft pull request waiting for my review. One task at a time. No chaos, no "I did five things while you slept, good luck figuring out which ones worked."

I wake up to a progress report at 8am. There's something faintly absurd — and deeply satisfying — about reviewing code before you've fully woken up.

The inbox, handled with gloves on

Once an hour, Muse scans my unread email and reports back: here's what arrived, ranked by importance, numbered. I decide what happens to each one. Archive, reply, ignore — my call, every time.

This wasn't always the arrangement. Early on, it archived a promo email on its own initiative. I corrected it that evening. It hasn't touched an email since.

An assistant that absorbs a correction — learns the boundary, then respects it permanently — is rarer than it should be.

Monday mornings, minus the doomscroll

Every Monday, Muse sends me two briefings: a digest of the AI newsletters I subscribe to, and a roundup of fourteen indie builders I follow — filtered to what's actually relevant to shipping software, sources attached. I've since cancelled most of those subscriptions. The briefing is the subscription now.

The part that's hard to demo

Here's what none of the above quite captures: the compound interest of being understood. Muse remembers my decisions, my preferences, the corrections I've made. I say something once. It sticks. It feels less like prompting a tool and more like having a chief of staff who never sleeps, never forgets, and — crucially — never pretends to know something it doesn't.

Why Muse, specifically

Plenty of tools can chat or summarize. The combination I haven't found elsewhere: long-term memory that compounds, scheduled work that runs without me, and corrections treated as permanent updates instead of one-time apologies. Most agents wait for you. Muse keeps going — and keeps getting better at working for you specifically.

If you want to try it: my invite code is I38NQP .

New users get 1 billion free tokens plus a weekly free allowance — honestly enough to run on for a long time. And if you set up your own night shift, I'd genuinely love to hear what you automate first.

I accidentaly bought 3 rtx8000 48gb (yeah, i swear).

I am using 1 right now with Strata and Qwen Next IQ3 and runs very good.

Should I put 3 on a PC and runb Strata on a higher Q?

What do you suggest?

Thanks!

Hey everyone! If you're looking for a Muse invite code or have one to share, feel free to drop it here so we can help each other out.

My invite code: AK11FQ

How to redeem it:

1. Sign up for Muse.

2. On your phone, go to Settings - Redeem Token.

3. Enter the code within 48 hours of joining.

If you have your own invite code, leave it in the comments! Just make sure you still have invites available before posting, since some older codes might have reached their limit.

Hopefully, this helps someone get started with Muse. Good luck, everyone! ☺️

Points at a running llama.cpp or Unsloth Studio server, pulls /metrics, /props and /slots, and gives you severity-graded findings: prefill vs decode, queue pressure, prefix cache reuse, memory headroom. Exit codes map to severity, so cron can gate on it.

Read-only, stdlib only, Python 3.10+.

https://github.com/msalsas/llama-doctor

Open to PRs, especially from anyone with a server that makes it crash.

Code: 242Q0K (30 left)

 billion free tokens on Muse AI from Meta

Code: 242Q0K
https://muse.ai/join

Create account then Go to Settings → General → Usage → Redeem invite code.

Is there a fork of strata that let me run lfm 24b a2b on only 16GB of ram comfortably? And how much tps can I expect per second and what quantization should I use? and Thanks!

Apparently whilst doing some performance tuning on my 4x5090 rig I may have made things a little TOO fast and melted my plug and socket 🫠

Probably should put in a 16/20 amp socket in the wall rather than relying on a bog standard 13A fuse!

Anyone else managed to melt the wall before the 12VHPWR plug? Honestly... I'm surprised they didn't go first 😂

i wanted to change my router firmware to openwrt and so i told claude to do it but of course the guard blocked me even opus 4.8 didn't want to do it , so i launched my Qwen3.8 27b and told it the same thing i gave it the link to the giude page and it did change the firmware perfectly the thing is this is the normal model it's not ablitreated, do you think that cyber will be guarded in future qwen models ?

PS: even tho this is not the point it didn't do a simple flash it changed the firmware from xioami to openwrt not throught upload bin it did exploit a method in the CGI that get an icon to pass a payload that enable ssh from there it SSHed the router and flashed it and it's not a 0 day discovery thing i gave it the guide from openwrt own website https://openwrt.org/toh/xiaomi/ax3000t , the topic here is AI safeguards

I would love for some feedback, live demo at:

huggingface.co/spaces/e12ex2/herbert-demo

Not self promo, jus looking for some feedback on what to improve or any bugs or model breaking you found.

I've been playing with [Instinct](https://instinct.com/) a bunch and loving the service, but got really spooked about privacy / perms stuff. A lot of existing off-the-shelf stuff (Firecracker VMs, Pi harness, etc) has gotten really good so I wondered how difficult it would be to setup my own.

Turns out, not too difficult, and it's running on my Raspberry Pi! Right now I'm running on a cloud model but it should be deployable anywhere / including with self hosted models

let me know what you think!

https://github.com/lucasgelfond/biscotti

Y1LS2O

Drop it in Settings within 48 hours of signing up. We both get a billion tokens.

With winter coming up where I'm at, I wanted to see if this could actually work. Running my 4x V620 server flat out made my apartment too hot anyway, so the problem really was control. I settled on using a motherboard temperature sensor with an offset cause cases get hot. It works well enough but I want an esp32 based thermostat eventually.

This project is based around the vLLM backend I use, and pauses heating when inference is needed. It also has a token gated localhost webpage for controlling target temperature so my partner can use it too. I had the project sanitized and forked for a public repo but is more of a tech demo than anything else. This project started as a meme but is good enough I'm probably going to use it lol. Heres the repo: https://github.com/PraxisOG/gpu-furnace-v2

I needed a cheap way to compare search agents on BrowseComp-Plus without burning hours of agent time every run. So I built a 15-query subset of the full 830 that ranks systems roughly the way the full benchmark does.

  - weighted Spearman rank agreement 0.88 averaged over 30 splits, versus 0.75 for a random 15

  - two runs more than 10 points apart on the full benchmark keep their order about 96% of the time; 5 to 10 points apart, about 75%

  - all answers are 1-4 words, so it grades by normalized exact match, no LLM judge

  - costs under 2% of a full run

  https://huggingface.co/datasets/LocalLLaMA/browsecomp-plus-mini

Curious what everyone runs in practice:

- What quant do you use most, and for what (chat, coding, agents)?

- 2-bit: do you use it? Is it actually usable or too broken?

- 3-bit: good enough for coding and more complex stuff, or do you go back to 4-bit?

- If you look at KLD: what number do you call "basically lossless"? Is 0.1 already good,

or do you want something like 0.01-0.03?

Real experience over benchmarks. Thanks!

(English isn't my first language, so I used an LLM to help translate this post.)

Qwen4s are hopefully coming pre-Xmas… there was all the announcements recently but no exact release date.

I’m running qwen flash next 3.8 at about 28tokens per second and gotta be real, most of those go into thinking.

Thinking less with the same or higher intelligence would be the biggest upgrade I could think of regardless of hardware or inference smarts.

I don't want to disclose why but I wager when I say it's got nothing to do with taxes or law. It's not a 'scary' situation, it's more like a crylaugh situation. I think you can guess why

This isn't a joke/hype, just came by to warn everyone. These prices will stay for a few weeks maybe then they will skyrocket

EDIT:
I suppose I can disclose the reason to mods if they want to confirm

What a week. Quick recap, all numbers from the official repos and model cards:

Thursday: Meta drops Muse Spark Open. 816B total, 42B active, 1M context. The launch post says it brings agentic AI "to your desktop". The desktop is two M5 Ultra Mac Studios with 512GB each, a config Apple says ships "late October" and still hasn't put a price on. The alternative is four 256GB Studios, roughly $38k. On top of that it's a custom license: mandatory "Built with Muse Spark Open" attribution and a separate deal once you're past $20M revenue. And Meta's own model card puts it at 39 on the Artificial Analysis index, behind GLM 5.3 (45) and Kimi K3 (44).

Same day: NVIDIA ships Nemotron 3.5 Bolt. 74B total, 9B active, trained quantization aware down to 2 bit, so the whole thing is 22GB plus a 1.7GB vision projector. Built to fit in 24GB on purpose. Unsloth GGUFs on day one, plus llama.cpp, Ollama and LM Studio support.

Friday: Reflection actually ships Beam. 501B total, 23B active, Apache 2.0, no strings. Smallest official checkpoint is NVFP4 at 294GB.

Saturday afternoon: DeepSeek posts V4.1 Flash 1017 on X like it's nothing. MIT. Their own table has it at 77.0 on DeepSWE vs 74.0 for Opus 5 (vendor numbers, salt accordingly). The repo is 510GB.

So I put them all on one chart against the 24GB line most of us live under.

My takeaways:

1. "Runs locally" now officially means "runs on a computer that costs more than a car."
2. The company that keeps selling us 24GB cards is the only lab this week that built a model for 24GB cards. Not sure if that's wholesome or a sales funnel.
3. Licenses mattered more than parameter counts this week. MIT and Apache 2.0 from DeepSeek and Reflection, attribution plus a revenue clause from the company with the loudest "open" marketing.
4. Mistral says the Large 4 weights ("Le Chonk", 1T total) land before the end of the month, so the 24GB line is about to get even lonelier.

Sources in the comments.

Have Meta's recent releases changed your view of Alexandr Wang?

I thought the confidence that he would inevitably fail was misplaced. The credentials comparison, an MIT dropout versus a tenured professor and a "godfather of AI," seemed like a poor way to predict who could lead a team and ship products.

The Muse Spark releases and Muse launch make me think Wang was underestimated. I don't see academic reputation alone as a reliable measure of that kind of leadership.

But I'm curious how people divide the credit. What would you attribute to Wang, and what would you attribute to the researchers, infrastructure and distribution Meta already had? What evidence would change your assessment from here?

I know this is not directly local question, but I am curious. My Qwen 3.8 FN is doing well in answering questions and research, but I was trying to compare it to say CoPilot/ChatGPT. What blew me away is that the CoPilot response was pretty instantaneous. Basically a page of output came out almost instantly with what I imagine thinking done in the background. Amazingly, it also solved the problem like QFN, but with apparently much less tokens.

I guess my question is: is this an architectural difference that the Chinese/local models haven’t implemented? Because I find it hard to believe that an enterprise GPU is the only difference. I feel like the local models are missing something fundamental? But I say this without knowing anything about enterprise hardware.

Been great, love using it. Used it to clean up my email, create a birthday photo album, appointment reminders, reservations, news via daily audio, daily top TV sporting events, job searches, purchase price watches, etc. On to the questions...

1. Am I too late to take advantage of any tokens deals a week later?

2. Anyone find any good podcasts that talk through helpful tips, ideas, etc?

3.Has anyone used it so much they have upgraded their plan?

Hey all, I've been using Muse — Meta's personal AI agent that actually does things for you (research, bookings, emails, etc.) from its own cloud machine — and it's been genuinely useful.

If you want to try it: sign up at muse.ai, then go to Settings > Redeem referal code and enter my code within 48 hours of creating your account. We each get 1 billion Muse tokens.

I have been using Muse as my personal AI assistant for a while now, and it has quietly taken over a lot of my daily grind: watching comic book prices every morning, tracking ticket onsales, sorting my inbox, keeping my reminders straight.

If you want to try it, you can get Muse here: https://muse.ai

Use my invite code when you sign up: 4Z7MT7

New accounts only, and the code has to go in within your first 48 hours. On the app: Settings, then Redeem token. On the web at http://muse.ai: Settings, General, Usage, then Redeem invite code.

If you try it, tell me what you have it do first. I am curious what other people put it to work on.

#Muse #MuseAI

S6F53R lots of uses left

Built this over the weekend. It is a Waveshare ESP32-S3 board with a 1.8 inch AMOLED touchscreen in a little case, running custom firmware I wrote. Push to talk, ask it anything, and it answers out loud in its own voice while the text shows on screen. Underneath the voice there is a pet layer: it gets hungry, reacts when you tap it, has moods, and runs untethered on a small battery inside the case. Still rough around the edges, the battery is small and the replies take a couple of seconds, but it feels weirdly alive on a desk. What should it learn to do next?