Meta has introduced Muse Spark, a new closed-source large language model that now powers its Meta AI app and website. Previously developed under the codename Avocado, the model marks the first release from the company’s restructured AI efforts under Meta Superintelligence Labs.
Muse Spark brings native multimodality, meaning it can handle text, images, and other inputs without relying on separate systems. It also includes built-in reasoning capabilities, tool-use, visual chain-of-thought processing, and support for multi-agent orchestration. These features position it as Meta’s attempt to compete more directly with leading frontier models from OpenAI, Google, and others.
The model is now live for all users on meta.ai and in the updated Meta AI mobile app. A private API preview is also available to selected developers. Meta describes Muse Spark as the initial step on what it calls its “scaling ladder,” with further investments planned across research, training, and infrastructure, including its Hyperion data center.
In internal evaluations, the “Thinking” mode of Muse Spark shows competitive results against models such as Opus 4.6 Max, Gemini 3.1 ProHigh, GPT 5.4 Xhigh, and Grok 4.2 Reasoning, particularly in multimodal perception, health-related tasks, and agentic workflows. The company acknowledges remaining gaps in long-horizon planning and complex coding scenarios, areas where it continues to invest.
A third mode, currently in development and named “Contemplating,” will orchestrate multiple agents working in parallel. Early benchmark results for this unreleased mode include 58 percent on Humanity’s Last Exam and 38 percent on FrontierScience Research, aiming to match the most demanding reasoning setups from competitors.
The update also brings a visual refresh to the Meta AI interface, though the core shift is the replacement of earlier models with Muse Spark. While the new system improves on Meta’s previous offerings, real-world performance will ultimately be judged by everyday users across creative, practical, and technical tasks.
Meta’s move reflects the broader industry pattern of rapid model iteration and heavy infrastructure spending. Whether Muse Spark delivers a noticeable leap in daily interactions or simply narrows the gap with rivals remains to be seen once broader feedback accumulates. For now, it represents another incremental advance in the ongoing competition among consumer-facing AI assistants.
