Meta's Muse Wonder and Muse Wonder Full aren't just two more chatbots with a fresh coat of paint. They're the first models Meta built specifically to run AI agents that stay with you — across days, weeks, and app launches — instead of starting from a blank slate every time you open a chat window. If you've ever re-explained your entire situation to an assistant for the fifth time in one week, that's exactly the annoyance these two models are aimed at.
What happened
Meta has released Muse Wonder and Muse Wonder Full, and it calls them the most powerful models in the company's history. Both were designed for "personal and professional" AI agents — software that acts on your behalf instead of just answering one question at a time. In plain terms, that means three things: the model remembers you from one conversation to the next, it doesn't quit the moment it hits a wall, and it keeps working on a long job until that job is actually done.
Two models, two different doors
Muse Wonder is live right now through the Meta Model API and inside Muse Code, Meta's coding agent, and it's been available since October 12. Muse Wonder Full is the heavier model, and it's locked to Muse Code subscribers. One detail worth spelling out: Full isn't a "think harder" toggle sitting inside Wonder. It's a separate model entirely.
Internally, the project ran under the codename "Watermelon." Alexander Wang, describing the launch, didn't hedge: "These are true frontier models." Frontier is industry shorthand for pushing the outer edge of what anyone can do today — a claim worth noticing, since Meta isn't usually the first company to make it.
Where Muse Wonder sits next to Muse Spark
Muse Spark, the previous model in the family, was built to be light and fast: low delay, quick answers, the kind of model you reach for when you just want a reply. Wonder and Wonder Full are the bigger siblings — more muscle, aimed at genuinely hard work like programming. Meta ties the quality jump to larger base models plus changes in how they're pretrained and trained with reinforcement learning.
Those two terms, without the jargon: pretraining is the long study phase where a model soaks up enormous amounts of text and media, the way you absorb a language by living in a country for a year. Reinforcement learning is training by reward — much closer to teaching a dog with treats than to handing someone a textbook.
The specs, translated: the context window, meaning how much the model can hold in mind at once, is around 1 million tokens. A token is a chunk of text, often a word fragment, so a million tokens is a serious amount of material — think several novels at once. It can take in text, images, video, and audio. It puts out text only. And it was trained in Prometheus, Meta's gigawatt-scale data center in central Ohio. Wonder is also headed to the Muse app soon, starting with subscribers, where Meta says it will handle artifact generation and "processes that require more intelligence."
Active Memory: it remembers you between chats
This is the feature the whole release hangs on. Active Memory means the model carries important context about you from one conversation to the next and pulls those details back up when they're actually relevant to the task in front of it. Today's assistants mostly don't do this, so you end up re-teaching them your preferences, your constraints, and your circumstances in every new thread.
Picture a good human assistant. They remember your client's deadline, that your kid has a dentist appointment on Thursday, and that you hate being called before 10 a.m. Active Memory is Meta's attempt at that, in software form.
Muse researcher Kavya Kopparapu said the team put serious effort into remembering what matters to a user across days, weeks, and months — not just inside one conversation. Meta says that on MemoryBench, a benchmark built around long-term memory and continuous learning, Muse Wonder comes out ahead of Claude Fable, Claude Opus, and GPT-6 Astra. Those are Meta's own numbers, so apply the usual grain of salt.
Active Focus: it won't walk away mid-task
Active Focus is the persistence half of the pair. Rather than stopping at the first obstacle, the model is meant to try a different approach, route around the blockage, adapt when new information shows up, ask a clarifying question when something is vague, and check in with you before taking an important action.
Wang says the model was trained specifically not to get stuck on roadblocks and not to come back to you over every tiny detail. Meta tests this with FocusBench, which measures sustained effort on hard tasks, and SWE-Marathon, which is built around long software projects that run far past the length of a normal chat.
Where the numbers land
Meta compared its models against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1, and claims parity or better on agentic, coding, and professional benchmarks. On the published table — Muse Wonder Full versus Claude Opus 5.5 versus GPT-6 Astra:
• SWE-Marathon (long software builds): 57.5% / 46.3% / 42.5%
• DeepSWE v1.1 (long-horizon coding): Muse Wonder Full edges out both rivals — a real lead, but a narrow one.
• Harvey LAB-AA (legal tasks): Muse Wonder Full finishes clearly on top, with GPT-6 Astra second and Claude Opus 5.5 behind it.
• Personal Superintelligence Index (Meta's own metric): Muse Wonder Fullfirst, Claude Opus 5.5 second, GPT-6 Astra third.
Spot the pattern. The gaps are widest on the longest tasks. On DeepSWE the win is real but far from a rout — the three models land close enough that a bad day could flip the order. The legal benchmark is the eye-opener: this is the one test where Muse Wonder Full doesn't just nose ahead of the pack, it opens a margin the other two can't touch. And on Meta's own scorecard the ranking holds, which is Meta's point, delivered by Meta.
What it means for you
At home
Think about planning a kitchen renovation. You tell the assistant once about your budget, your landlord's rules, and the fact that your partner vetoes anything dark green. Three weeks later, when you open a new chat to compare countertop quotes, you don't start over — it already knows the constraints. The same logic applies to a messy job like "compare these twelve insurance plans and tell me which three actually cover water damage." That's a multi-step errand you can hand off and walk away from, which is very different from babysitting a chatbot through twelve rounds of copy-paste.
At work
Most knowledge work is long work: a contract review, a research memo, a quarterly report that gets revised over two weeks. Today you act as the memory for your AI — pasting in old context, correcting the same mistake twice a day. Active Memory flips that. You brief it once, and the next session picks up where the last one stopped. For anyone juggling five projects, the time saved isn't in the writing. It's in the remembering.
If you run a business
This is where agent-style models start to look like a hire rather than a toy. A support backlog, a sequence of follow-up emails, a competitor-tracking routine — jobs that used to die halfway because the model gave up or forgot the brief. Small teams get the biggest lift here, because a one-person shop can now run work that used to need a coordinator to keep it moving.
For students
A literature review or a semester-long project is exactly the shape these models are built for. You can hand over a reading list, ask for summaries as you go, and come back a week later to a thread that still knows which sources you've already covered and which argument you were building. The caveat is obvious and worth saying out loud: remembering your context is not the same as being right, so keep checking the work.
For creators
If you write a series, run a channel, or build a brand voice, you know the pain of re-explaining your tone in every new draft. A model that holds onto your style guide, your character list, and your no-go phrases across sessions turns fifty small prompts into one long collaboration. It also means less of that weird drift where episode twelve sounds like it was written by a stranger.
A new income angle
The practical shift: you can now sell outcomes instead of hours. A bookkeeping routine that runs weekly, a legal-prep workflow for small firms, a tutoring assistant that tracks each student's weak spots over months — these are small products a solo operator can actually maintain. You're no longer selling "I'll prompt a chatbot for you." You're selling a process that keeps running.
How to try it right now
1. Start free. Before you spend anything, test whether a memory-style workflow even fits your job. Free AI tools like the ones collected at MyKreaTool are a good, zero-cost way to run a two-part task across two sessions and see what breaks.
2. If you have Meta Model API access, call Muse Wonder directly. It's been available through the API since October 12, so there's no waiting list to join.
3. If you code, open Muse Code and select Muse Wonder as your model. Subscribers can switch over to Muse Wonder Full, which is the more powerful of the two and only lives inside that subscription.
4. Watch for the Muse app. Wonder is coming there soon, subscribers first, and it'll be the easiest entry point once it lands.
5. Test it properly. Give it a task that spans sessions — something with a real deadline and a couple of interruptions. That's the only honest way to judge Active Memory and Active Focus for yourself.
Upsides and what changes
The biggest change is the unit of work. Until now, the chat was the unit: one question, one answer, start again. With memory and persistence built in, the unit becomes the project. That single shift is what makes long-running agents feel plausible instead of like a demo.
The second change is competitive pressure. Meta is claiming the top of the market on long-horizon coding and legal tasks, with results out ahead of GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. Even if you discount vendor benchmarks by half, that's a fight that pushes every other lab to ship better memory, and you get the benefit.
Limitations
Be honest about the fine print. Every number here comes from Meta's own testing, on Meta's own benchmarks, including the Personal Superintelligence Index, which is Meta's own scorecard — nobody grades their own homework for free. Several of the published figures are so close together that they're better read as a tie than a victory. Muse Wonder Full sits behind a Muse Code subscription, and the announcement doesn't mention pricing at all, so the real cost is still a question mark. The output is text only; it can read images, video, and audio, but it can't hand you back a picture or a voiceover. A 1-million-token window is also not the same thing as flawless memory — holding a lot in mind and using it correctly are two different skills, and handing an AI months of your personal context raises fair questions about what gets stored and for how long. Finally, benchmarks describe average tasks, and your workflow is not average; the only test that matters is the one you run on your own work.
Conclusion
Meta's pitch with Muse Wonder and Muse Wonder Full is simple: stop treating AI as a search box with a personality, and start treating it as a colleague who was in the room last week. The memory and persistence features are the real story — not the benchmark table, which is Meta grading Meta, and where the closest results are close enough to argue about.
One action for today: pick the single task you're sickest of re-explaining to an AI, write it down, and run it twice — once now, once tomorrow — to see whether the context survives. If it does, you've just found your first real agent job.
Comments 0