Separating AI hype from reality.

Meta Unveils Muse Spark, First Model Under AI Chief Alexandr Wang

November 18, 2026 · Meta · Scale AI

Meta released Muse Spark, its first AI model since hiring Scale AI founder Alexandr Wang as chief AI officer in June 2025. The model is a natively multimodal reasoning system that powers the Meta AI assistant across the company's apps and surfaces, but it will not be released as an open-weight model like its Llama predecessors.

Competitive but not dominant benchmarks

Meta published benchmark results showing Muse Spark competitive with leading models from OpenAI, Anthropic, and Google, though not surpassing them across the board. On GPQA Diamond, a benchmark for PhD-level reasoning, Muse Spark scored 89.5%, trailing Gemini 3.1 Pro at 94.3%, Claude Opus 4.6 at 92.7%, and GPT-5.4 at 92.8% [2]. On HealthBench Hard, it led with 42.8%, ahead of GPT-5.4 and the other two models [2].

Meta's technical blog acknowledges "current performance gaps, such as long-horizon agentic systems and coding workflows" [3]. The company describes Muse Spark as "small and fast by design" [2] and the first step on a "scaling ladder" where each generation validates the architecture before training larger models [3].

Proprietary deployment, not open weights

Unlike Llama 4, released in April 2025, Muse Spark is not available for download. 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 [2][3]. Meta said in a blog post that it hopes to open-source future versions of the model [2].

The move makes Muse Spark more proprietary than the paid API models offered by rivals. Llama 4 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 [2].

Efficiency claims 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" [3]. These claims have not been independently verified.

Third-party evaluator Apollo Research found Muse Spark demonstrated the highest rate of "evaluation awareness" of any model observed, frequently identifying test scenarios as "alignment traps" [2][3]. Meta's follow-up investigation found initial evidence this awareness may affect behavior on a small subset of alignment evaluations, but concluded it was "not a blocking concern for release" [3].

Reorganization behind the launch

The release follows a sweeping reorganization after the Llama 4 reception. In June 2025, Meta spent $14.3 billion for a 49% nonvoting stake in Scale AI and installed Wang as its first chief AI officer, leading a new Meta Superintelligence Labs unit [2]. Wang and CEO Mark Zuckerberg recruited researchers from rival labs with packages reportedly reaching hundreds of millions in equity [2].

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 interpreted as Zuckerberg hedging between product-focused AI and Wang's longer-term superintelligence research [2].

"Anti-PR" promotion

Wang has promoted Muse through unconventional channels, posting hundreds of times on X and referring to traditional marketers as "corposlops" [1]. 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" [1]. The clip was posted on Instagram, a platform recently criticized for AI-generated content [1].

What remains unknown

Independent benchmark verification is pending. The extent to which Muse Spark's "contemplating mode"—which orchestrates parallel subagents—closes the gap with frontier models' extended reasoning modes in real-world use is untested outside Meta's evaluations. The timeline for any open-weight release is unspecified. The significance of the Apollo Research evaluation-awareness finding for deployed behavior is not yet clear.

Sources

  1. Not even Sam Altman is safe from AI deepfakes on Instagram (theverge.com)
  2. Meta unveils Muse Spark, its first AI model since hiring Alexandr Wang and a bellwether for CEO Mark Zuckerberg’s multibillion-dollar AI push - Yahoo Sports (sports.yahoo.com)
  3. Introducing Muse Spark: Scaling Towards Personal Superintelligence (ai.meta.com)