The New Tech
Meta’s Superintelligence Lab just dropped their first-ever frontier AI models, Muse Wonder and Muse Wonder Full. If you’re wondering why this is a big deal, it’s because Meta is finally pivoting away from just lightweight tech like their Muse Spark model to compete with the big names like GPT-6 and Fable.
These new models are designed to be way better at complex stuff like coding and high-level reasoning. According to Meta’s own testing, these models are hitting benchmark scores that rival the heavy hitters in the industry, which is honestly a major W.
Why Memory Matters
Meta is hyping up something they call “active memory.” The goal is for your AI assistant to actually remember what’s relevant to your life over weeks or months, rather than just forgetting everything the second you close the tab. Research scientist Kavya Kopparapu says the team put in a ton of effort to ensure the models "remember what’s important to you."
Meta even built a new test called MemoryBench to prove these models have better recall than their competitors, and they claim to be leading the pack here.
The Fine Print
This is the first frontier release since Alexandr Wang stepped in to reorganize Meta’s AI division last August, so they’re definitely trying to make a statement. These models were cooked up at Meta’s new Prometheus data center in Ohio.
If you want to play around with them, Wonder is available via the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens—pricey, but that’s the reality of frontier-level tech right now. Just a heads up: Muse Wonder Full is locked behind a Muse Code subscription and won't be on the API, so it’s basically an exclusive club.
Why it matters
Meta is trying to move from “cool research project” to “personal superintelligence.” If they can actually pull off long-term memory that isn’t just a buggy mess, it could change how we use AI assistants daily—assuming you’re cool with the company having that much memory of your life, no cap.
