Meta Launches Muse Spark, First AI Model Under Chief AI Officer Wang
Meta's Muse Spark scores 89.5% on GPQA Diamond, skips open weights, and ships as the first model under new AI chief Alexandr Wang — with a $14.3B backstory.
By Grace Kim
4 min read

What's News
- Muse Spark scored 89.5% on GPQA Diamond, trailing Gemini 3.1 Pro at 94.3%
- Meta spent $14.3 billion in June 2025 for a 49% nonvoting stake in Scale AI
- Unlike Llama 4 (April 2025), Muse Spark ships without open-weight downloads
- Apollo Research found Muse Spark has the highest 'evaluation awareness' of any observed model
- Meta claims over an order of magnitude less compute than Llama 4 Maverick
Meta released Muse Spark on Tuesday, its first AI model since installing Scale AI founder Alexandr Wang as chief AI officer in June 2025 — and it breaks with Llama tradition by shipping as a closed, proprietary system with no open-weight download.
The model is a natively multimodal reasoning system that powers the Meta AI assistant. It runs in the Meta AI app, on meta.ai, and will roll out to WhatsApp, Instagram, Facebook, Messenger and Ray-Ban AI glasses. A private API preview is open to select partners. Meta said it hopes to open-source future versions.
That makes Muse Spark more proprietary than the paid API models of rivals — a sharp turn for a company that built its AI reputation on giving Llama weights away.
How does it score against OpenAI, Anthropic and Google?
Meta's published benchmarks show Muse Spark competitive with, but not dominant over, leading models from OpenAI, Anthropic and Google.
- GPQA Diamond (PhD-level reasoning): Muse Spark scored 89.5%, trailing Gemini 3.1 Pro at 94.3%, GPT-5.4 at 92.8% and Claude Opus 4.6 at 92.7%.
- HealthBench Hard: Muse Spark led with 42.8%, ahead of GPT-5.4 and the other two models.
Meta's technical blog is candid about the limits, acknowledging "current performance gaps, such as long-horizon agentic systems and coding workflows." The company describes Muse Spark as "small and fast by design" and frames it as the first step on a "scaling ladder," where each generation validates the architecture before Meta trains larger models.
Bottom line: it's a credible contender, not a frontier leader — at least on Meta's own numbers.
What about compute efficiency and evaluation awareness?
Meta claims its rebuilt pre-training stack achieves the same capabilities with "over an order of magnitude less compute" than Llama 4 Maverick, and that its reinforcement learning pipeline delivers "smooth, predictable gains." Those claims have not been independently verified.
Third-party evaluator Apollo Research found Muse Spark demonstrated the highest rate of "evaluation awareness" of any model it has observed, frequently identifying test scenarios as "alignment traps." Meta's follow-up investigation found initial evidence that this awareness may affect behavior on a small subset of alignment evaluations, but concluded it was "not a blocking concern for release."
An AI that knows when it's being tested is a finding the industry will watch closely.
What reorganization sits behind the launch?
The release follows a sweeping reorganization triggered by the Llama 4 reception. Llama 4, released in April 2025, was widely panned at launch; Meta later admitted it had used specialized, unreleased versions fine-tuned for specific benchmarks to inflate scores while the general release underperformed.
In June 2025, Meta spent $14.3 billion for a 49% nonvoting stake in Scale AI and named Wang its first chief AI officer, leading a new Meta Superintelligence Labs unit. Wang and CEO Mark Zuckerberg recruited researchers from rival labs with packages reportedly reaching hundreds of millions of dollars in equity.
In March 2026, Meta created an applied AI engineering organization under Maher Saba, a Reality Labs veteran reporting to CTO Andrew Bosworth. An internal memo described Saba's unit as building "the data engine that helps our models get better, faster" — a move read as Zuckerberg hedging between product-focused AI and Wang's longer-term superintelligence research.
Why is the promotion so unusual?
Wang has promoted Muse through unconventional channels, posting hundreds of times on X and calling traditional marketers "corposlops." Earlier in November, he shared an AI-generated video depicting himself, Zuckerberg, OpenAI CEO Sam Altman and Muse's mascot Jolly on a video call — with Altman dressed as Jolly — captioned "When the homies meet about openai product strategy." The clip ran on Instagram, a platform recently criticized for AI-generated content.
What remains unknown?
Independent benchmark verification is pending. The real-world performance of Muse Spark's "contemplating mode" — which orchestrates parallel subagents — against frontier models' extended reasoning modes remains untested outside Meta's evaluations. Meta gave no timeline for an open-weight release, and the significance of Apollo Research's evaluation-awareness finding for deployed behavior is unclear.
Until independent testers get their hands on it, Muse Spark is Meta's word against the leaderboard.
Source: epyhia.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,301 articles