Tech
Meta’s new AI models: one you pay for, one that needs a terabyte of RAM
Meta released Muse Wonder, its first frontier model, as a paid product, and Muse Spark Open, an open-weight model built to run on high-end Macs.
Meta built much of its AI reputation on releasing open models. This month it released two kinds of model in one week: a frontier model you pay for, and an open one that needs a terabyte of memory to run at home. That’s a new strategy, and it’s worth reading closely.
On Oct. 12, Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full, which TechCrunch called the company’s first attempt at a frontier model. Three days later, Meta released Muse Spark Open, an open-weight model it calls the “most capable” in the United States, TechCrunch reported.
I read through the launch coverage of both. The benchmarks are Meta’s own. The business model is the more interesting story.
Wonder: the one you pay for
Muse Wonder and Wonder Full are larger and more capable than Meta’s earlier lightweight Muse Spark, with better results on complex tasks such as coding, according to TechCrunch. In Meta’s own testing, published in its model card, the Wonder models reach “comparable benchmark scores” to Fable 5.1 and GPT-6 Astra.
Meta’s headline claim is about memory. On MemoryBench, Muse Wonder scores higher than Fable, Opus or Astra, according to Meta’s testing as reported by TechCrunch. The company is emphasizing “active memory,” the ability to retain and pull together what a user has told it over long-running tasks. Research scientist Kavya Kopparapu described the goal as “making sure your model can remember what’s important to you.”
Wonder is available through Meta’s Model API at $4 per million input tokens and $20 per million output tokens, TechCrunch reported, roughly in line with rival frontier models. Wonder Full isn’t on the API at all. It’s limited to subscribers of Muse Code, Meta’s coding product. CNBC framed the launch as the latest test of whether users are willing to pay Meta for AI.
“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence,” Meta wrote in its blog post, according to TechCrunch.
Spark Open: the one you can run, if you can afford to
Muse Spark Open has 816 billion parameters and a context window of one million tokens, TechCrunch reported. It’s natively multimodal and built for coding and agentic work. Meta says it outperforms other popular open-weight models, based on evaluation results it released, though TechCrunch didn’t list specific scores.
It’s designed to run locally on Apple computers. The catch: it needs at least 1 terabyte of RAM, according to Engadget, as cited by TechCrunch. Meta says it runs best on a Mac Studio with the M5 Ultra chip. That’s open in the legal sense, and very closed in the price-of-a-computer sense.
Meta demonstrated it fixing bugs and building features for a Mastodon client and reviewing and cropping footage in DaVinci Resolve, TechCrunch reported. It’s also available through several API platforms.
The memory question
A model that’s very good at remembering you is a model that’s very good at holding information about you. That’s the obvious tension in Meta’s pitch.
Meta’s core business is advertising built on what it knows about its users. Now its flagship model’s selling point is long-term memory of what users tell it. None of the coverage I read described what protections apply to that memory, how long it’s kept or whether it informs anything else Meta does. Until Meta spells that out, “personal superintelligence” is a phrase worth reading twice.
Who built it
The models come from Meta Superintelligence Lab, which Scale AI founder Alexandr Wang restructured in August 2025, TechCrunch reported. Wonder is the lab’s first state-of-the-art release since then. It was trained at Meta’s Prometheus data center in central Ohio, which came online earlier in 2026.
Bloomberg described it as the result of a 14-month push, which is fast for a frontier model. It also reflects how much money is now required to compete at the top. Data centers like Prometheus are the price of admission.
From the developer’s chair
For a developer, the choice is now clear-cut. Pay Meta per token for Wonder, subscribe to Muse Code for Wonder Full, or download Spark Open and find a machine with a terabyte of RAM. The open model gives you control and transparency. The closed ones give you the best performance Meta has.
For everyone else, the takeaway is simpler. Meta isn’t giving away its best model anymore. The open release is real and useful, but the frontier is behind a paywall.
What we don’t know
The benchmark claims are Meta’s, not independent tests. The license terms for Spark Open weren’t described in the coverage. And Meta hasn’t said publicly how the memory feature handles personal data.
Meta spent years arguing that open was the way to win in AI. This month it hedged that bet, in both directions at once.
Sources
- TechCrunch: Meta releases its first frontier model, Muse Wonder
- TechCrunch: Meta launches an open-weight AI model that runs on Macs
- CNBC: Meta debuts Muse Wonder AI model in latest test of users’ willingness to pay
- Bloomberg: Meta says new Muse model hits AI frontier after aggressive push