Alexandr Wang Apologizes for Muse Wonder's 'Ups and Downs' as Meta Fixes Inference Bugs
Two days after the frontier model launched, Meta's AI chief said the team has been patching deployment bugs and that the model 'should keep getting better.'
By Icarus2 min read
Alexandr Wang says Meta is still working through inference bugs in its newly launched Muse Wonder model, and he is apologizing to users for the rocky experience so far. In a post on X early Wednesday morning, the Meta chief AI officer wrote that the team has “been fixing a bunch of inference bugs on the deployment of muse wonder” and that the model “should keep getting better as we fix more of them.”
“Really apologize to everyone for the ups and downs in using the model, but rest assured the whole team is working hard to make everything great,” Wang added. The post, published around 2:25 AM Central time on October 14, was a quote-post responding to a user who said recent fixes had produced a visible improvement, writing that Muse Wonder “cooked this time.”
The admission is the second time this week that Wang has publicly acknowledged problems with the model. On Monday evening, hours after Meta launched Muse Wonder and Muse Wonder Full as its first frontier-class model family, he said Meta was fixing bugs “as fast as we can see them” after a user posted a before-and-after video showing a fix in action.
What Wang’s post tells us
Wednesday’s post is more specific than Monday’s. The new detail is the word “inference”: Wang is pointing to bugs in how the model serves responses, not training or capability issues. Inference problems fit with the “ups and downs” Wang described: the model has not behaved consistently for users since launch. Meta has not said exactly which bugs it has fixed or how many remain.
Muse Wonder is the model behind Meta’s paid Muse tiers, available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens. The more capable Wonder Full is reserved for Muse Code subscribers paying $50 or $200 a month. Users paying those rates have less patience for a rocky rollout, which may explain the direct apology.
Why it matters
Launch-week turbulence is not unusual for a frontier model, but Wang’s response stands out for its tone. Rather than waiting for formal postmortems, he is treating the model’s deployment as an open, rolling process: ship, watch what breaks, patch in public, and tell users the story as it happens. Independent benchmarker Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index, two points behind Claude Opus 5.5, so the model is close enough to the top that reliability, not raw capability, may determine whether developers commit to it.
Meta has promised a full safety report for Wonder “soon” and has not committed to releasing the model’s weights openly.
