Meta launches Muse Spark, a natively multimodal reasoning model and its first proprietary AI from Meta Superintelligence Labs

Meta announced Muse Spark on April 8, 2026, the first model from Meta Superintelligence Labs — a new AI research and product organization built with former Scale AI co-founder Alexandr Wang, in whom Meta invested $14.3 billion for a 49% stake. Muse Spark marks Meta's first step away from its open-source Llama lineage toward a proprietary frontier model, available immediately at meta.ai and through a private API preview.

What's new

Muse Spark is described by Meta as "a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration." The model was rebuilt from scratch — Meta's team spent nine months redesigning its entire AI stack before shipping Muse Spark as the first in the new Muse series.

Key capabilities and performance:

• Multimodal by design: Handles vision, language, and tool use in a unified model, not as a post-hoc add-on
• Contemplating mode: A multi-agent reasoning mode that "orchestrates multiple agents that reason in parallel" — designed for complex, multi-step reasoning tasks
• Efficiency: "We can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick" — a major internal efficiency gain
• Benchmarks: 58% on Humanity's Last Exam (HLE) and 38% on FrontierScience Research in Contemplating mode
• Availability: Live at meta.ai and the Meta AI app; rolling out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses; private API preview open to select users

The model targets competitive performance in "multimodal perception, reasoning, health, and agentic tasks" — putting it in direct competition with GPT-5 class models and Gemini 3.5.

Context

Muse Spark is a sharp departure from Meta's AI strategy of the last three years. Since 2023, Meta built its AI reputation almost entirely on Llama — a series of open-weights models that became the foundation of the open-source AI ecosystem. Llama 4 Maverick, the most recent Llama model, was itself a strong open model.

The pivot to a proprietary model began with the formation of Meta Superintelligence Labs, the hiring of Alexandr Wang, and the $14.3 billion Scale AI deal. Wang's Scale AI provides the data labeling infrastructure that underpins frontier model training — Meta acquired deep insider access to that data pipeline along with Wang's leadership.

The Muse series represents Meta's answer to an increasingly clear pattern in the industry: open-weight models are catching up to closed frontier models, but the very frontier — where the most capable systems live — is still driven by proprietary development and training infrastructure.

Why it matters

Meta entering the proprietary frontier model race changes the competitive landscape. The company has the compute scale of a hyperscaler, the user surface of the world's largest social networks, and a new data infrastructure advantage through Scale AI. Muse Spark's multimodal-first, agentic design also signals where Meta sees the next battleground: not chat completion benchmarks, but multi-step reasoning, agent orchestration, and real-world task completion.

For developers, the private API preview suggests a commercial model API is coming. If Muse Spark's claimed efficiency gains hold — "over an order of magnitude less compute than Llama 4 Maverick" for the same capability level — it could also be offered at price points that disrupt existing API pricing tiers.