Meta's Muse Spark: Alexandr Wang's First AI Model Reshapes the Frontier Race with Multi-Agent Reasoning and Medical AI Dominance

Meta has fired a major shot in the AI race. On April 8, 2026, the company unveiled Muse Spark, the first model from its newly formed Muse series, built from the ground up by Meta Superintelligence Labs (MSL) under the leadership of former Scale AI CEO Alexandr Wang. Codenamed "Avocado" during its nine-month development cycle, Muse Spark represents a fundamental departure from Meta's Llama lineage — and a bold statement that the social media giant is no longer content playing catch-up in the frontier AI arena.

A New Chapter: From Llama to Muse

For years, Meta's AI identity was synonymous with Llama, its family of open-source language models that democratized access to large-scale AI. But when Alexandr Wang joined Meta in mid-2025 after his surprise departure from Scale AI, it signaled a dramatic strategic shift. Wang was given a mandate: build a model that could compete head-to-head with the best from OpenAI, Google DeepMind, and Anthropic — and do it without the constraints of the open-source philosophy that defined Llama.

The result is Muse Spark, Meta's first closed-source frontier model, and it has already begun reshaping conversations about what's possible in AI.

What Makes Muse Spark Different?

Muse Spark is a natively multimodal reasoning model — meaning it doesn't just process text, images, and audio separately, but fuses them during inference in a unified architecture. Its headline feature is “Contemplating mode,” a novel multi-agent parallel reasoning system where internal specialized agents collaborate to break down complex problems before producing a final response.

Illustration of a neural network architecture. Image: Wikimedia Commons (CC BY-SA 3.0)

Think of it as an internal “council of experts.” When you ask Muse Spark a difficult medical question, one internal agent might focus on clinical data, another on pharmacological interactions, and a third on patient context — all working simultaneously before synthesizing an answer.

Benchmark Battle

Meta claims Muse Spark achieves state-of-the-art performance across several key benchmarks:

• MMLU-Pro: 91.2% (surpassing GPT-5's reported 89.8%)
• MedQA (USMLE): 96.1% (highest score ever by an AI system)
• GPQA Diamond: 78.4% (competitive with Gemini Ultra 2)
• HumanEval+: 94.7% on coding tasks
• MathVista: 71.8%

The medical AI performance is particularly noteworthy. A 96.1% score on USMLE-style questions significantly outperforms the average human physician score of approximately 60-70%. Meta has announced partnerships with Mayo Clinic and Johns Hopkins to explore clinical applications.

The Closed-Source Controversy

Meta, long the champion of open-source AI, has made Muse Spark entirely proprietary. Wang defended the decision: “Open science remains core to Meta's mission, but frontier safety demands frontier responsibility.”

The AI research community is divided. Yann LeCun, Meta's Chief AI Scientist, has been notably quiet. Critics have called the move a betrayal of Meta's open principles, while supporters argue the closed approach enables more robust safety guardrails.

What This Means for the AI Landscape

• Five serious frontier contenders: OpenAI, Google DeepMind, Anthropic, xAI, and Meta's MSL.
• Medical AI has a new leader: The USMLE performance could accelerate regulatory conversations.
• Multi-agent architectures going mainstream: Contemplating mode validates the approach.
• Open vs. closed debate intensifies: Meta's pivot may shift industry strategies.

Looking Ahead

Muse Spark will be available through a new API starting Q3 2026, with enterprise pricing undercutting OpenAI and Google by approximately 30%. Consumer access through WhatsApp, Instagram, and Facebook is expected by late 2026.

Meta has also teased “Muse Flame,” a next-generation model targeting AGI benchmarks. Under Alexandr Wang's leadership, Meta's AI ambitions have entered a new and formidable chapter.