Meta’s new “small and fast” AI model, Muse Spark, is an acknowledgement that as enterprises scale AI systems beyond millions of users and for use on a greater variety of devices, they must make things more efficient and more application-specific.

Muse Spark now powers the Meta AI assistant on the web and in the Meta AI app, and the company plans to roll it out across WhatsApp, Instagram, Facebook, Messenger, and the company’s smart glasses. It will also offer select partners access to the underlying technology through an API, initially as a private preview.  “We hope to open-source future versions of the model,” it said in a blog post announcing Muse Spark.

While Meta did not disclose the model’s size or much about its architecture, it described Muse Spark as being capable of balancing capability with speed.

That positioning, even without explicit enterprise deployment guidance, aligns with priorities CIOs and developers are increasingly grappling with as they move generative AI from pilots to production, focusing on efficiency, responsiveness, and seamless integration into user-facing software.

The model’s other capabilities, including support for multimodal inputs, multiple reasoning modes, and parallel sub-agents for complex queries, could help enterprises build faster, task-focused AI for customer support, automation, and internal copilots without relying on heavier models.

Meta said it has worked with physicians to improve responses to common health-related questions, underscoring the model’s applicability across a range of use cases, including reasoning tasks in science, math, and healthcare.

It said it had conducted extensive pre-deployment safety evaluations, with particular attention to higher-risk domains such as health and scientific reasoning. The company also touted said it had made improvements in refusal behavior and response reliability, aimed at reducing harmful or unsupported outputs.

It published the results of 20 AI benchmarks for Muse Spark, positioning it as competitive in several areas while not claiming across-the-board leadership. In particular, it highlighted strong performance on health-related assessments, reflecting its focus on improving responses in that domain through targeted training and evaluation.

The model also scored well on multimodal and reasoning-oriented benchmarks, sometimes a little ahead of rivals such as Claude Opus 4.6, Gemini 3.1 Pro, GPT 5.4 or Grok 4.2, sometimes a little behind.

Meta frames the model as part of a broader roadmap, with future models expected to extend capabilities further, suggesting a staged approach rather than a single model designed to lead on all benchmarks.