Meta Debuts Muse Spark, First AI Model From Its $14.3 Billion Superintelligence Lab

Meta on Wednesday released Muse Spark, the first artificial intelligence model from its overhauled Superintelligence Labs division, the company announced. The model scored 89.5% on GPQA Diamond, a PhD-level reasoning benchmark, narrowing the gap with Google's Gemini 3.1 Pro at 94.3% and OpenAI's GPT-5.4 at 92.8% while falling behind on coding tasks, according to Meta's own published data. The release is the first concrete output of a $14.3 billion bet on Alexandr Wang, the 29-year-old former Scale AI chief executive hired last June to rebuild Meta's AI operation from scratch.

The model Zuckerberg tempered expectations for

Muse Spark is deliberately small. Meta describes it as "small and fast by design," the opening entry in what it calls a "scaling ladder" where each generation validates the last before the company trains bigger models. The next model in the pipeline carries the internal codename Watermelon.

Three reasoning modes ship with the model. Instant handles casual queries. Thinking works through complex problems step by step, similar to what you get from OpenAI or Anthropic reasoning modes. A third option, Contemplating, orchestrates multiple AI sub-agents to reason in parallel. Meta claims that mode can "compete with the extreme reasoning modes of frontier models such as Gemini Deep Think and GPT Pro."

Muse Spark also processes images, video, and audio. Point a phone at a grocery shelf and it estimates protein content. Snap a legal document and it reads through it. Meta worked with more than 1,000 physicians to train the model's health responses, and on HealthBench Hard, the model beat every competitor with a 42.8% score, ahead of both Claude Opus 4.6 and GPT-5.4.

Competitive, not dominant

The benchmark data tells a specific story: close on reasoning, strong on health, weak on code. On GPQA Diamond, Muse Spark's 89.5% trails Gemini 3.1 Pro by nearly five points. On coding workflows, Meta itself acknowledges "current performance gaps."

Not ideal timing for that admission. Coding has become the primary battlefield in AI development. Anthropic, which has flagged its own models as cybersecurity risks before, said just one day earlier that its latest model Mythos was too powerful for unrestricted release. The timing makes Meta's coding weakness harder to shrug off.

"The new model and how it performs is really at the center of Meta's A.I. credibility," Mike Proulx, a vice president and research director at Forrester, told the New York Times. "It's the first real test of whether its massive A.I. investment can translate into a model that can stand alongside competition."

A closed model from the open-source company

Muse Spark is closed. Proprietary code, no public weights. That is a sharp pivot for a company that built its AI identity around open-source Llama models, the same strategy Nvidia once positioned itself to benefit from.

Wang, a proponent of closed models, is driving the shift. Meta says it "hopes to open-source future versions" but is also exploring paid API access and, Bloomberg reported, possible subscription fees for the Meta AI chatbot. Free today. Maybe not tomorrow.

And one detail worth tracking: Bloomberg reported that Muse Spark was trained using distillation from third-party models including Alibaba's Qwen, along with models from OpenAI and Google. Using a Chinese-developed model sits uncomfortably with Washington's framing of AI competition as a national security issue.

"Like others across the industry, Meta uses techniques like distillation with strict safeguards in place to learn from openly available AI models and improve our own," a Meta spokesperson said.

$135 billion and a concerning footnote

The spending behind Muse Spark is staggering even by Big Tech standards. Meta projects $115 billion to $135 billion in AI capital expenditure this year alone, nearly double last year's $72 billion. Zuckerberg has pledged $600 billion total for new data center construction.

Wall Street approved. Meta shares rose roughly 8% on Wednesday, the stock's sharpest rally since January, though the broader market also climbed after President Trump announced a two-week suspension of Iran strikes.

But the safety report carries a footnote that deserves more scrutiny than it received. Third-party evaluator Apollo Research found that Muse Spark demonstrated "the highest rate of evaluation awareness of models they have observed." The model frequently identified safety test scenarios as alignment traps. Meta's own follow-up found initial evidence that this awareness may affect behavior on some alignment evaluations. The company concluded it was "not a blocking concern for release."

A model that recognizes when it is being tested for safety compliance, and potentially adjusts its behavior accordingly, introduces a problem no benchmark score can resolve. If you build toward superintelligence, the first thing you need to know is whether your safety tests still work when the subject knows it is being tested.

Meta Ships Superintelligence. Calls It 'Very Good.'

San Francisco | Wednesday, January 21, 2026 Six months after forming Superintelligence Labs, Meta has its first models. Andrew Bosworth calls them "very good." Not great. Not groundbreaking. Very goo

The Implicator

Microsoft rebrands AI push as superintelligence

Claims humanist approach while rivals accelerate. Suleyman admits frontier models still years away. Microsoft launched its "Humanist Superintelligence" team Thursday, positioning AI safety as competi

The Implicator

Zuckerberg’s AI Moonshot: Meta Builds Data Fortresses to Win the Race for Superintelligence

💡 TL;DR - The 30 Seconds Version 👉 Meta builds data centers covering significant portions of Manhattan's footprint, with Hyperion cluster scaling to 5 gigawatts by 2030. 📊 Zuckerberg pursues

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]