Alexandr Wang stood on a stage at Meta's Menlo Park campus nine months ago, freshly poached from Scale AI with a $14.3 billion deal that made Silicon Valley's jaw drop. Wednesday, his team finally shipped the result: Muse Spark, Meta's homegrown AI model that it says "significantly narrows the performance gap" with OpenAI, Anthropic, and Google.

Narrows the gap. Not closes it.

A Meta executive told Axios something remarkable in a launch-day briefing — that Muse Spark "doesn't mark a new state of the art." For $14.3 billion, they got a model that's competitive "at certain tasks" like multimodal understanding and health information processing, while acknowledging an outright gap in coding. The model accepts voice, text, and image inputs but only produces text output. In April 2026, when competitors are shipping natively multimodal systems with 2-million-token context windows, that feels like arriving at a Formula 1 race in a really nice sedan.

Here's what Meta actually bought with that $14.3 billion: a 49% stake in Scale AI and Alexandr Wang's brain. Wang, who built Scale AI into a data-labeling empire serving nearly every major AI lab, now leads Meta Superintelligence Labs. His team built Muse Spark — code-named Avocado — from the ground up over nine months, and it represents a significant upgrade over Meta's Llama 4 models.

The model rolls out immediately in the Meta AI app and Meta.ai website, with plans to expand across Facebook, Instagram, and WhatsApp. It features a fast mode for casual queries, several reasoning modes, and a "shopping mode" that combines LLMs with user interest and behavior data. That last part is where Meta's real play becomes obvious — this isn't about building the best AI. It's about building the best AI that can sell you things.

Meanwhile, the competitive landscape isn't waiting around. Anthropic this week detailed Mythos, a model so powerful that its initial release is limited to a handful of tech companies for cybersecurity defense. OpenAI is finishing a new model code-named Spud that reportedly represents another significant leap. Google just shipped Gemini 3.1 Ultra with a 2-million-token context window working natively across text, image, audio, and video.

My Opinion

I'll be blunt: $14.3 billion for "not state of the art" is a sentence that should make Meta shareholders uncomfortable. That's not pocket change — that's more than the GDP of several small countries, spent to acquire one very talented person and a chunk of his old company, to produce a model that Meta itself admits can't beat what's already on the market.

Here's what bugs me about this whole thing. Meta spent years positioning itself as the open-source champion of AI. Llama was the people's model. Now Muse Spark launches as proprietary, with vague "hopes to open-source future versions." That pivot tells you everything about where the industry is headed — when the models get good enough to matter commercially, the open-source evangelism dries up real fast.

But the shopping mode reveal is the honest part. Meta doesn't need the best AI model in the world. It needs an AI model that's good enough to keep 3.9 billion monthly users inside its apps, clicking on ads, buying things through Instagram shops. The $14.3 billion wasn't an investment in artificial general intelligence. It was an investment in artificial general commerce. And on those terms, it might actually pay off — just don't pretend this is about pushing the frontier of human knowledge.

Author: Yahor Kamarou (Mark) / www.humai.blog / 12 Apr 2026