Meta Muse Spark - The Open-Source Giant Goes Closed

Apr 12, 2026 · #meta #muse spark #artificial intelligence #open source #alexandr wang #llama #ai models

For years, Meta was the company that gave AI away for free. The Llama series became a symbol of openness. Now? Closed model, closed code, closed doors.

Abstract closed doors with digital code in the background

Meta just released Muse Spark - the first model from its new Meta Superintelligence Labs division. It’s led by Alexandr Wang, who Meta brought in for a staggering $14 billion (previously CEO of Scale AI). Here’s the catch - the model is completely closed-source.

No weights to download. No code to inspect. Nothing.

The company that spent three years championing open AI just did a full 180.

How does it actually perform?

Muse Spark landed in 4th place on the Artificial Analysis Intelligence Index v4.0, scoring 52 points. Ahead of it: Gemini 3.1 Pro (57), GPT-5.4 (57), and Claude Opus 4.6 (53).

Fourth place isn’t bad. But it’s not exactly groundbreaking either.

There are two areas where Muse Spark does lead, though. HealthBench Hard - difficult medical questions - score of 42.8. And Humanity’s Last Exam, essentially an expert-level knowledge test - 50.2%. Nobody else comes close on those.

”Thought compression” - what does that even mean?

Meta claims Muse Spark uses something called “thought compression.” In practice, this means the model needs 10x less compute than Llama 4 Maverick to reach its answers.

Sounds impressive. But that’s Meta’s claim, not an independent test. We’re taking their word for it right now.

The model also has built-in multimodal understanding - text, image, video processed together. That’s not unique in the industry anymore, but Meta’s planning to deploy it at a scale nobody else can match.

Where will people actually see it?

Muse Spark is rolling out across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses. That’s over 3 billion users. No other AI company has that kind of distribution reach.

And that’s probably the real reason behind going closed-source. If you’ve got 3 billion people you can serve AI directly - why hand the model to your competitors for free?

Is this the end of open AI from Meta?

That’s the $14 billion question. Literally.

Meta says it “hopes to” open-source future versions. But “hopes to” and “plans to” aren’t the same sentence. Llama still exists and it’s still open. But the flagship, newest model? Locked down.

Alexandr Wang built Scale AI on the philosophy that data and models should be accessible. Now he’s building a closed model on Zuckerberg’s dime. Did he change his mind, or did Meta change it for him?

My take

I don’t think this is the end of open AI from Meta. It’s more of a signal that Meta’s now playing both sides. Llama stays open - because it attracts developers and builds an ecosystem. Muse Spark stays closed - because it protects competitive advantage.

The problem is that this undercuts Meta’s strongest argument in the AI wars. “We give it away for free” was the line that drew people to their side. Now that line comes with an asterisk and fine print.

And the results? Fourth place is decent, but for $14 billion I’d expect more. Wang’s just getting started, so the real test comes with the next versions.

AITU #04 - more on the channel.