Published: April 9, 2026
Meta has released Muse Spark, the first large language model developed by Meta Superintelligence Labs. The model is designed to be small and fast while delivering competitive reasoning, multimodal perception, and health-related capabilities. It currently powers the Meta AI assistant and will roll out across Instagram, WhatsApp, Facebook, Messenger, and Ray-Ban Meta AI glasses in the coming weeks.
What is Muse Spark
Muse Spark is the inaugural model in a new “Muse” series, developed entirely within Meta Superintelligence Labs under the leadership of Alexandr Wang, Meta’s chief AI officer who joined the company nine months ago. The lab rebuilt Meta’s AI stack from the ground up, taking a different approach from the company’s previous Llama series.
The model accepts voice, text, and image inputs but produces text-only output. Meta describes its development philosophy as “a deliberate and scientific approach to model scaling,” where each generation validates and builds on the previous one before increasing scale. This first release is intentionally compact, prioritizing speed and efficiency over raw parameter count.
Technical capabilities and efficiency
The most notable technical claim around Muse Spark is its efficiency. According to Meta, the model achieves its reasoning capabilities using over an order of magnitude less compute than Llama 4 Maverick, the company’s previous mid-size flagship model. This efficiency gain is attributed to a technique Meta calls “thought compression.”
Muse Spark handles complex questions in science, math, and health. Its multimodal perception enables strong image understanding, including visual coding capabilities. The model can also launch multiple subagents simultaneously for more comprehensive answers to complex queries. Meta has indicated that a “Contemplating” mode for tackling more difficult problems is planned for a future update.
How it compares to competitors
Meta executives have stated that Muse Spark does not represent a new state of the art across all categories. The model is competitive with leading models from other labs in certain areas, particularly multimodal understanding and health information processing. However, Meta acknowledges that a gap remains in other areas, most notably coding, compared to models already available from competitors like OpenAI and Anthropic.
The positioning of Muse Spark as a compact, efficient model suggests Meta is taking a different strategic path with this new series. Rather than competing purely on benchmark performance, the company appears to be optimizing for speed and deployment across its consumer products, which serve billions of users globally.
Availability and what comes next
Muse Spark is currently live in the Meta AI app and on meta.ai. It will be rolled out to WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses in the coming weeks. A private preview via API is available for select partners, with broader paid API access expected at a later date.
Meta has signaled its intention to open-source future versions of the Muse model series, though Muse Spark itself launches as a proprietary model. This marks a shift from the company’s Llama series, which was released with open weights. The relationship between the Muse and Llama product lines going forward remains to be clarified.
For more information, see Meta’s official announcement and the Meta AI technical blog post.
