Mark Zuckerberg just played his most expensive hand yet. Meta unveiled Muse Spark, the first AI model from its Superintelligence Labs — the division built on a $14.3 billion Scale AI investment and a talent raid that cost hundreds of millions in individual engineer packages. After more than a year in the wilderness following the embarrassing Llama 4 launch, Meta says it’s back.

But here’s the twist nobody expected: the company that championed open-source AI just went proprietary.

The Llama 4 Hangover

This doesn’t happen in a vacuum. Llama 4 wasn’t just underwhelming — Meta got caught manipulating benchmark results with specialized fine-tuned versions while the actual release model fell flat. Developers bailed. The open-source community that had rallied around Llama felt burned. Meta’s AI credibility cratered while OpenAI, Anthropic, and Google sprinted ahead.

Zuckerberg’s response was scorched earth. He brought in Alexandr Wang — formerly Scale AI’s CEO — to run a brand-new division called Meta Superintelligence Labs. The internally code-named “Avocado” project spent nine months “rebuilding our AI stack from the ground up.” That’s an extraordinary admission from a company that positioned itself as the AI industry’s scrappy open-source champion.

Competitive, Not Dominant

So how good is Muse Spark? Good enough to matter. Not good enough to lead.

Artificial Analysis scored it at 52 on its Intelligence Index — top five globally. That slots it behind GPT-5.4 and Gemini 3.1 Pro (both at 57), just behind Claude Opus 4.6 (53), but dramatically ahead of Meta’s previous embarrassments. Llama 4 Maverick and Scout scored 18 and 13 respectively.

The breakdown tells a nuanced story:

  • GPQA Diamond (PhD-level reasoning): 89.5% — trailing Gemini 3.1 Pro (94.3%), Claude Opus 4.6 (92.7%), and GPT-5.4 (92.8%)
  • HealthBench Hard (medical reasoning): 42.8% — beating every rival model, including all three leaders
  • Coding and abstract reasoning: Still lagging behind the top three

That health score is the headline. Meta collaborated with over 1,000 physicians to curate medical training data, and it shows. If you’re building health-adjacent AI products, this could be a genuine differentiator.

Wang acknowledged the gaps openly: “There are certainly rough edges we will polish over time.” Credit where it’s due — that’s a dramatic improvement over the Llama 4 benchmark manipulation fiasco.

The Open-Source Community Just Got Ghosted

Muse Spark is closed source. Not “limited release” closed — it’s more locked down than models from OpenAI and Anthropic. You can use it through the Meta AI app and meta.ai, eventually through WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban smart glasses. Select partners get API access via “private preview.” That’s it.

Meta — the company that made its name in AI by giving Llama away for free — just built a walled garden.

Zuckerberg offered the classic politician’s pivot: “Looking ahead, we plan to release increasingly advanced models… including new open source models.” But that’s a promise about future models, not this one. The strategic logic isn’t subtle. OpenAI and Anthropic’s combined valuations exceed $1 trillion. Meta’s open-source generosity didn’t translate into AI revenue. You don’t capture market value by giving your best work away.

Still, it stings. The open-source AI community just lost its biggest corporate patron.

3.5 Billion Users Is the Real Moat

Here’s where Meta’s strategy gets genuinely interesting. Unlike OpenAI or Anthropic, Meta doesn’t need to acquire users — it already has 3.5 billion of them.

Muse Spark is designed for practical, everyday tasks: estimating calories from a photo, visualizing furniture in your room, shopping recommendations embedded in the Meta AI chatbot. These aren’t flashy demos that win Twitter clout. They’re the kind of features that drive daily engagement at massive scale.

The model also ships with Contemplating Mode — Meta’s reasoning answer to Google’s Deep Think and OpenAI’s GPT Pro. It spins up multiple agents simultaneously to reason through complex problems in parallel. It’s natively multimodal across text, images, audio, and video, with tool use and multi-agent orchestration baked in.

The play isn’t to build the smartest model. It’s to build one that’s smart enough and deploy it to more users than any competitor can reach.

The $135 Billion Question

Meta’s AI capex for 2026 is projected at $115–135 billion — nearly double last year. Muse Spark needs to justify that number.

Zuckerberg tempered expectations on the January earnings call: “I think the team’s first models will be good but, more importantly, will show the rapid trajectory that we’re on.” Translation: don’t judge us on Muse Spark alone, judge us on the slope of improvement.

That’s either visionary patience or an expensive coping mechanism.

A Five-Horse Race Now

The AI model market in April 2026 is more crowded than ever. Each company has carved out a distinct identity:

  • OpenAI: The incumbent frontrunner, now exploring advertising revenue
  • Anthropic: The safety-focused lab with careful, limited releases
  • Google: The infrastructure giant with Gemini woven through search and cloud
  • Meta: The distribution king betting on personal AI for billions of users

Muse Spark doesn’t need to be the smartest model in the room. It just needs to be good enough to be useful to 3.5 billion people. Based on these benchmarks, it might actually be.

Whether “good enough” justifies $135 billion in capex — that’s the trillion-dollar question that will define whether Meta’s AI pivot is remembered as a masterstroke or the most expensive mistake in tech history.