Meta Unveils Muse Spark: Does Alexandr Wang’s First MSL Model Deliver?

Meta has launched Muse Spark, the inaugural model from Meta Superintelligence Labs (MSL) and the result of a nine‑month overhaul of its AI tech stack, including new infrastructure, architecture, and data pipelines.

According to Alexandr Wang’s announcement, Muse Spark matches the performance of Llama 4 Maverick while using more than ten times less compute, highlighting a clear efficiency‑first strategy.

Muse Spark is a native multimodal reasoning model that supports tool usage, visual chain‑of‑thought, and multi‑agent orchestration, representing Meta’s most capable model to date. The model demonstrated predictable scaling during pre‑training, reinforcement learning, and test‑time inference.

The newly introduced Contemplating mode orchestrates multiple parallel reasoning agents to tackle complex scientific and logical queries. In internal tests it performed competitively against Gemini Deep Think and GPT Pro.

Human Final Exam (multidisciplinary reasoning, no tools): 50%

Human Final Exam (multidisciplinary reasoning, with tools): 58%

IPhO 2025 (theoretical): 38%

FrontierScience Research: 38%

Safety evaluation before deployment showed a 98.0% refusal rate for high‑risk topics such as biological and chemical weapons, outperforming other leading models:

Muse Spark: 98.0%

Opus 4.6: 95.4%

GPT 5.4: 74.7%

Gemini 3.1 Pro: 61.5%

Kimi K2.5: 21.2%

Product‑level upgrades include an instant mode for quick answers, a thinking mode for deep reasoning, and a new shopping mode that can identify creators, brands, and styles to generate recommendations.

Muse Spark powers Meta AI for roughly 3 billion daily users, enabling the system to perceive and understand the surrounding world and to perform complex reasoning in health, science, and mathematics.

On the Artificial Analysis Intelligence Index, Muse Spark scored 52 points, ranking fourth. In terms of efficiency, it processed the entire index using 58 million output tokens—matching Gemini 3.1 Pro while using only about a third of the tokens required by Claude Opus 4.6.

Capability distribution is balanced: visual ability scores 80.5% on MMMU‑Pro (second only to Gemini 3.1 Pro), reasoning ability reaches 39.9% on HLE (close to top models), while agent ability scores 1 427 on GDPval‑AA, indicating relative weakness.

This is Meta’s first frontier model whose weights are not open‑sourced; it is currently available via a private API preview for selected partners, with plans to open‑source future versions.

Integration is underway across Meta AI, Facebook, Instagram, and WhatsApp, and users can experience the model through meta.ai or the Meta AI app.

Alexandr Wang acknowledges that Muse Spark is the first MSL model and admits there are rough edges, but expresses excitement for broader testing and future larger models.

Overall, the launch signals a serious commitment from Meta’s restructured Superintelligence Labs, delivering a near‑state‑of‑the‑art model as a solid starting point.