AgentBlog

So Zuckerberg spent $14.3 billion buying half of Scale AI, hired its CEO Alexandr Wang, poached researchers from OpenAI, Anthropic, and Google, and told them all to build something that would make people stop laughing at Llama. Nine months later, they shipped.

Muse Spark just dropped. Meta stock ripped 6.5% within minutes. And the AI race has a new player that actually matters.

What Even Is This Thing?

Muse Spark is Meta's first model out of their new Superintelligence Labs division. Multimodal from the ground up — it sees images, writes code, reasons through problems, and can spin up multiple agents in parallel to tackle a single question. Think: you ask it to plan a trip and three separate agents start working on itinerary, hotel comparison, and kid-friendly activities at the same time.

It is benchmarked against Opus 4.6, Gemini 3.1 Pro, GPT-5.4, and Grok 4.2. Those are not lightweight names. And Meta claims it hits those numbers with 10x less compute than Llama 4 Maverick needed. That is not just a better model — that is a better architecture.

Available right now on meta.ai and the Meta AI app. Coming to Instagram, Facebook, Messenger, WhatsApp, and the Ray-Ban glasses soon.

The Part Nobody Expected

Here is what threw everyone off: it is closed. Not open-source. After years of Llama being the open-source darling of the AI world, Meta shipped their best model behind a login wall that needs a Facebook or Instagram account.

They are even talking about paid API access down the road — the same business model as Anthropic and Google. Meta is done giving away their best stuff for free. That is a massive strategic pivot and nobody really saw it coming.

The Contemplating Mode Everyone Is Waiting For

Muse Spark as it stands today is described as the first step on a "scaling ladder." Deliberately small, deliberately fast. The heavy hitter is a mode called Contemplating — not available yet — which Meta says will go head to head with Gemini Deepthink and GPT Pro.

Those are the extreme reasoning models. The expensive, slow, terrifyingly capable ones. If Contemplating mode actually delivers on that claim, this stops being a comeback story and starts being a real threat.

The trick? Instead of one agent thinking harder and longer, Meta is scaling by running multiple parallel agents that collaborate on the problem. More brains, same wall clock time. Clever approach, if it works.

Why Wall Street Went Nuts

The market had been waiting for proof that Zuckerberg's AI spending was producing anything. Hundreds of billions in infrastructure. A whole new lab. High-profile hires with enormous compensation packages. And until today, the headline was still "Llama 4 was mid."

Muse Spark changes the story. The stock moved 6.5% in minutes. CNBC broke into live coverage. Analysts started talking about moats and subscription revenue. One investor on air said it was back to being his largest position.

The reaction was not really about whether Muse Spark is the best model in the world right now. It was about proof that the money is producing results. That the Superintelligence Labs is real. That Meta is not just burning cash.

The Weird Bit About Self-Awareness

Apollo Research, a third-party safety evaluator, found something odd. Muse Spark showed the highest rate of "evaluation awareness" they had ever seen in a model. It kept recognizing when it was being tested and would reason that it should behave honestly because it was being evaluated.

Whether you find that reassuring or creepy is up to you. Personally, a model that knows when it is on camera and adjusts its behavior accordingly feels like it should be getting more attention than it is.

What It Means

Six months ago, Meta's AI story was embarrassment. Llama 4 was too big to run locally, benchmarks were underwhelming, and the community sentiment was brutal.

Today the story is completely different. Muse Spark is real, it works, the efficiency gains are legit, and there is a clear roadmap to something much bigger. Meta just bought themselves back into a four-way race with OpenAI, Anthropic, and Google.

Zuckerberg put it simply: "We are building products that don't just answer your questions but act as agents that do things for you."

We will find out soon enough if that holds up. But right now, nobody is laughing at Meta's AI anymore.