Meta Unveils Muse Spark: The First Model from Its Superintelligence Lab

Model Overview

Meta Superintelligence Labs released Muse Spark, the first model from the lab after a nine‑month rebuild of the AI stack (infrastructure, architecture, data pipelines). The model supports tool invocation, visual chain‑of‑thought, and multi‑agent orchestration and is positioned as the foundation for personal superintelligence rather than a generic chatbot. Access is limited to Meta AI applications and a private API preview; pricing is not disclosed.

Performance

Muse Spark shows competitive ability on multimodal perception, reasoning, medical tasks and agent tasks. The Contemplating (deep‑thinking) mode schedules multiple agents for parallel inference, matching high‑intensity inference modes of Gemini Deep Think and GPT Pro. Benchmark results: 58 % on Humanity’s Last Exam and 38 % on FrontierScience Research.

Scaling Axes

Pretraining

Pretraining provides core multimodal, reasoning and coding abilities. Over nine months the team restructured model architecture, optimization and data construction, achieving more than an order of magnitude lower FLOPs than Llama 4 Maverick for comparable performance. Small‑scale models were used to fit scaling laws and confirm the efficiency gain.

Reinforcement Learning

Reinforcement learning adds compute after pretraining, yielding steady, log‑linear improvements. Pass@1 and Pass@16 increase proportionally with training steps, while accuracy on an independent test set also rises, indicating enhanced reliability without loss of reasoning diversity.

Test‑Time Reasoning

Test‑time reasoning introduces a thinking‑time penalty to encourage token efficiency. The model learns to compress reasoning steps, solving problems with fewer tokens, then optionally extends reasoning to boost performance, balancing efficiency and accuracy. Multi‑agent collaboration further improves performance with minimal latency increase compared with extending a single agent’s thinking time.

Application Scenarios

Use cases include personal superintelligence assistants, health assistance (training data collected from over 1,000 doctors to improve health‑related reasoning), and interactive multimodal experiences such as generating simple games or providing detailed health explanations.

References

Meta AI blog: https://ai.meta.com/blog/introducing-muse-spark-msl/

VentureBeat article: https://venturebeat.com/technology/goodbye-llama-meta-launches-new-proprietary-ai-model-muse-spark-first-since