The end of Llama? Meta launches Muse Spark
Meta AI has unveiled Muse Spark, the first AI model from its newly formed Meta Superintelligence Labs (MSL). This marks a significant overhaul of the company’s AI strategy. Designed as a foundation for personal superintelligence, Muse Spark is a natively multimodal reasoning model that combines tool use, visual chain-of-thought, and multi-agent orchestration. It aims to deliver advanced performance across perception, health, coding, and agentic tasks.
Muse Spark emphasizes efficiency and scalability through improved pretraining techniques, reinforcement learning, and test-time reasoning. Pretraining enhancements allow the model to extract more capability per unit of compute, achieving performance comparable to previous larger models (such as Llama 4) while using an order of magnitude less compute. Reinforcement learning contributes predictable gains in reliability and reasoning diversity. Test-time reasoning incorporates extended thinking time and multi-agent orchestration to optimize token usage and reduce latency.
A standout feature of Muse Spark is Contemplating mode, which enables multiple agents to reason in parallel rather than extending the thinking time of a single model. This boosts performance on complex tasks, with Meta reporting significant gains on challenging benchmarks, including up to 58% on Humanity’s Last Exam and 38% on FrontierScience Research in this mode. The model also demonstrates strong reasoning compression, distilling complex thought into fewer tokens without major accuracy loss. These chains can later be expanded for enhanced overall performance.
Muse Spark is purpose-built for deep integration across Meta’s ecosystem. It currently powers the Meta AI app and meta.ai website, with rollout planned in the coming weeks to WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses. Its native multimodal perception supports image and visual data analysis, enabling applications ranging from health guidance and interactive coding to game creation. Health-related reasoning draws on curated, physician-verified data for detailed and factual responses to common queries.
The model further incorporates social and contextual awareness. It can reference relevant information from content shared across Meta’s platforms, delivering richer, more informed responses. This enhances experiences in areas such as shopping, travel planning, and content discovery.
Meta conducted extensive safety evaluations under its Advanced AI Scaling Framework. These assessments indicate that Muse Spark operates within safe boundaries across key frontier risk categories (including cybersecurity, biological/chemical risks, and loss of control). Safety measures encompass pretraining data filtering, post-training alignment, and system-level guardrails to support responsible deployment, even in dual-use scientific domains.
The release of Muse Spark represents a clear strategic shift for Meta – moving beyond the open-source Llama family toward a new series of models aimed at personal superintelligence. The company plans to expand the model’s capabilities iteratively, with private API previews available to select partners and future embedding of its intelligence into additional products and devices.