alt Hacker NewsMuse Wonder
Comments
I don't use Meta much, but was curious to test this out... went to sign into my Meta account (Quest only) from the Muse Code site https://dev.meta.ai/products/muse-code and had other people's profiles listed in my login auth -- looks like developers / people I have no relation to.
Looks like Meta may be exposing your profile and email to other people right now. Accounts settings are inaccessible to me to deactivate my account. Getting spammed with login code requests that I'm not generating.
Anyway, be wary.
Tried it out, asking it for a broad suite of real dev work:
- UI component polish on a React FE
- General review of a complex cross-cutting Node backend changes
- Minor changes to a Rust server with reviews by Codex
Well, my first impression is that it's still a step below Astra/Opus5.5/Fable here. Intelligence is there I think, but 'taste' and 'judgment' is not there yet. Personality seems okay but that doesn't help when it got stuck into a loop by being fixated on the review comments (on iteration 7 of review->fix last I checked) without taking a step back and reconsidering, and this is on a relatively simple change.
Limits seems okay-ish on the $50 plan but not too generous I think.
Image->HTML test for Muse Wonder. It's much worse than Opus 5.5. IMO it's closer to Sonnet or Haiku for this test.
Muse Wonder (high): https://html.non.io/lcars-muse-wonoder/
Opus 5.5 (high): https://html.non.io/lcars-opus-5.5/
Original design: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
I was expecting much more given the benchmark numbers it was putting out. I suspect its vision model isn't mature yet, and that it's been heavily optimized for text at this point in time.
Some pretty decent pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Here's Muse Spark 1.2 in August for comparison: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And since I seem to have missed doing pelicans for Muse 1.3 in September, I just ran those for a comparison: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... (1.3 seems to like animating them.)
I'm sorry zuck, i wasn't familiar with your game
2nd place model on aa according to them, losing only to opus 5.5 is really really not bad
Pricing is available here: https://dev.meta.ai/models/muse-wonder
Input -- $4.00
Cached input -- $0.20
Output -- $20.00
Edit: Upon closer inspection, the benchmark scores for "Muse Wonder" on the main marketing page are actually for "Muse Wonder Full," which is only available on the higher-tier subscription plans. This pricing is for "Muse Wonder."
If this model is not benchmaxxed, I fear it's over for anthropic, openai etc. They are not unique and there will probably be many others with similar models very soon. It'll become a commodity and the price will be driven to the ground.
When they released Muse Spark on september 2nd, they promised it would be open weights (https://xxcancel.com/finkd/status/2095232032896946311). We're still waiting for that...
Interestingly unlike other Meta models this one seems to take a pretty big jump in pricing, as opposed to the prior ones, where they typically came out at a highly discounted price compared to every other model.
Maybe its a signal they think they've now cracked into the frontier and are willing to actually price similarly vs just trying to get users with discounted but maybe slightly worse results?
$4/$20 in vs out making it higher than Sol and on par with Opus.
It's hard to try and compare these benchmarks against real world use cases. As an experiment I let Meta Muse make all major decisions on a weekend trip to Seattle ~2 weeks ago. I think this would have been using Muse Spark 1.3, which is equivalent to GPT 6 - Sol.
On the plus side it told me taking the light rail from SEA into the University district was faster and cheaper than Uber.
On the downside it would recommend restaurants that were out of business, checking out Salmon swimming through locks out of season, got the opening of Beecher's Cheese in Pikes Place wrong, and fumble it's way through most simple planning decisions. I think my favorite was when it asked me what flight I have tomorrow so it could help plan when to head to SEA. It sent me a very panicked message that the flight was leaving today! Very soon! I had to remind it this is a daily flight and I am booked on the next day before it checked and apologized.
For things that are verifiable, and code tends to be that way, these models can do all sorts of things. But I find it hard to reconcile Meta's goal of having a personal assistant bot with something that is unable to do the simplest tasks.
I trust Meta with my codes as much as I trust it with my nudes
This is still an entire generation behind in real-world performance. I'd say it's even behind ds4.1f in the limited testing I have done so far. It actually reminds me of 5.5 haiku more than anything, but the prompt intelligence is as bad as mimo models.
OpenAI halved the value of their pro subscription, literally. It kicks off on Oct 28. And simultaneously is now two steps back on the frontier? It's a confusing moment for that company as they try to IPO.
I have a feeling "intelligence too cheap to meter" is going to haunt/mock this industry. Recall that only a year and a half ago Llama 4 flopped so hard Meta rebooted their development efforts. Desperately wish more information was available publicly about proprietary model architectures and training regimes. Are they running Manhattan project style compartmentalization? Otherwise, are these details, and the economic implications, an open secret in their circles given the amount of crossover between companies? It'd be of great interest to us mere mortals that have to deal with the consequences of how this turns out.
Apparently “the models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year”
https://www.nbc4i.com/news/local-news/new-albany/meet-promet...
The benchmarks I saw posted to X are impressive, which makes them hard to believe.
I recall that Meta made massive investments into compute infrastructure that put their capacity near the top of all competitors. If all of this is true, it looks like ownership of compute is a significant differentiator in performance.
I wish I could benchmark Muse Wonder Full, because Muse Wonder isn't much better than Muse Spark 1.3[1]. I created a brand new programming language designed to confuse LLMs[2] and created a model benchmark out of it.
A lot of the tasks were failed because they exceeded the $5 budget though, so maybe it would have gone on to complete them, it just required a lot of funds to do so. Still, that's the benchmark, and it's not better than MiMo V2.6 Pro which is pretty disappointing.
Hopefully it's benchmaxxing. I am terrified of the idea of Zuck leading the frontier of intelligence.
As someone who got a Muse Code sub last month, I find this announcement a bit disheartening because it makes me think the lower tier subs are going to get rug pulled. I'm on the $15 plan for a month now and it's value with Muse Spark 1.3 Max-contributor beats everything else by a long shot. Two somewhat small prompts with Muse Wonder Max-contributor used 4% of my weekly quota and I quickly switched back. I was thinking about getting the $50 plan and I'm actually less likely to do that right now.
Would be great if Meta drops the weights, just to throw off Ant and OAI a bit
Just checked their subscription is still North America only, I guess I will keep myself and my company with claude...
These regional lockdowns are making me feel funky, I feel like I should just use open models if others start playing at this game as well.
I was/am fine with Chinese companies offering cheaper subscriptions in China because ofc. But this is just weird, you could always charge me more let me try the damn thing, atleast it's own open router it seems like.
Initial impressions, interesting, it generated a pretty ok result for my prompt need to run a larger task but don't want to burn money so will wait for someone to burn their subscription and try running some opus 5.5 comparisons.
The only time I'll ever touch a Meta LLM is if it's powering the dystopian doom bot that is plunging its murder spike through the base of my skull.
In the context of Meta trying to get non-technical people to embrace Muse agents to handle personal affairs, their internal metric showing a ~3% violation rate for mishandling personal information seems unacceptably high.
Not that interesting unless it's made open weights. Frontier lasts weeks, inference is a commodity.
spinning up a GPU and as many cores as possible?
mousewheel repurposed?
background follow your mouse?
neutral tones?
well then YOU are viewing a frontier model announcement!
So, Wonder (and not Wonder Full) will "will soon power experiences in" their assistant app. But no rollout details .
Then, Full is meant for API usage and "Muse Code" only.
It's getting confusing what having "Muse" means.
Reminded me of when George Hotz said only Zuck has “the button”. [0]
The Mishandling Personal Information benchmark is non-sensical: that's the very definition of giving your personal data to Meta
While some are afraid that one rouge AI system may take over the world (or cause doom is some fashion), I think one of the best safety mechanisms to prevent this is having many competing and adversarial labs that are extremely close to each other. So it's good to see that Meta is not showing any intention of leaving the frontier.
Does it weight tokens on their ability to elicit negative emotions in users in order to drive engagement?
It's interesting that in its marketing Meta compares Wonder to the closed frontier models but not the open (Chinese) models. By its own assessment, Meta could easily show how far ahead its own model is against the rest of the pack of open models but they chose not to. Why?
They added fancy annimations and effects to this specific webpage, which are unnecessary and distracting and lead to bad reading experience. I already suffer astigmatism and floaters. I hate it when I move my cursor and these gray blocks and characters flash and flicker.
I found it interesting in the example given of a user looking at house listings and the agent contacting the letting agent on it's own on behalf of the user.
Like the good ol' days of viruses and anti-viruses. I wonder if there's two sides that will play out; many agents across many users doing things at scale for their users, thereby creating an admin burden for someone else. However because the task is known there's a framework for how to solve the task, might Meta then provide an agent on the other side for the business owner to use.
Looks good, and at least it is available unlike Gemini 4 Argon
Acquihires starting to finally pay off it seems. Bring on the competition
Have I been living under a rock, or this feels out of nowhere from Meta?
Seems they’ve just messed around with cheap and open weight models for years.
Then Muse appears, followed by full suite of frontier models…
RL team hill-climbing this to 2nd AA place and this fast is crazy
How does it work on day-to-day workflows? does it translates well?
its three big companies at the frontier now? openai, anthropic and now meta.
why is google still not frontier?
I noticed Muse Wonder model on OpenRouter is age gated to 18+
oh boy, you can imagine the shit eating grin on the face of whoever came up with the name "wonder full".
If I'm reading this right, then at the moment, for certain tasks, this is the best model in the world.
A new company on the frontier. Not close, but actually on it. Wow this is big news.
This is amazing. I can’t wait benchmark to come out $ per performance
Apparently the benchmark-winning (behind Opus 5.5) Muse Wonder Full is only currently available to their subscribers.
Only Muse Wonder is available in the API.
Muse Glimmer, Muse Spark, Muse Wonder?
What's next, Sun Shine?
Meta's RL scaling seems to be in full force
impressive benchmark numbers… now how to actually use this without Zucking up my machine is a consideration.
I personally have many problems with Meta, but imo the more companies close to the frontier, the better.
If only one company were calling all the shots once they hit recursive self-improvement, it would lead to extreme concentration of power, and their values being enforced across the entire industry. With 3-4 companies at the frontier, there is more urge to cooperate to avoid a catastrophic race condition.