Meta Launches Muse Spark, Its First AI Model Under Alexandr Wang and Meta Superintelligence Labs

Meta Launches Muse Spark, Its First AI Model Under Alexandr Wang

Meta released Muse Spark on April 8, 2026, the first major model from Meta Superintelligence Labs. The lab, led by chief AI officer Alexandr Wang, rebuilt Meta's AI stack from the ground up over the past nine months.

What Muse Spark Does

Muse Spark is a natively multimodal reasoning model. It accepts voice, text, and image inputs and produces text output. It supports tool use, visual chain-of-thought reasoning, and multi-agent orchestration.

Meta also released Contemplating mode, which runs multiple agents that reason in parallel. This lets Muse Spark compete with the reasoning modes of models like Gemini Deep Think and GPT Pro.

Performance

Muse Spark scores 52 on the Artificial Analysis Intelligence Index, placing it behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Meta said that improved training techniques and rebuilt infrastructure allowed it to create smaller models that match its older midsize Llama 4 variant at one-tenth the compute cost.

Availability

Muse Spark is available on the web and the Meta AI app. Unlike Meta's earlier open-weight Llama models, Muse Spark is currently a closed model used primarily within Meta's own products.

Why It Matters

This release marks a strategic shift for Meta. The company moved from open-weight distribution to a closed model approach, signaling it views AI model quality as a competitive advantage worth protecting. The speed of development, nine months from lab founding to model release, shows the pace of frontier AI research continues to accelerate.