Meta Deploys Muse Spark On Smart Glasses

According to Meta, the company has begun rolling out Muse Spark across the Meta AI experience and is deploying it on smart-glasses hardware, reaching Ray-Ban Meta and Oakley Meta in the US and Canada, with Meta Ray-Ban Display slated for summer (ai.meta.com; About.fb). UploadVR and Built In report that Muse Spark is the first model released by Meta Superintelligence Labs, the group led by Alexandr Wang, and cite Meta's claim that the new model matches the performance of Llama 4 Maverick while using 10x less compute (UploadVR; Built In). UploadVR also reports leaderboard-style benchmark numbers, listing Muse Spark at 52 versus 57 for Gemini 3.1 Pro and 60 to 61 for newer competitor models.

Per Meta's public materials, Muse Spark is described as a natively multimodal reasoning model that processes text, images, and speech, with support for visual chain-of-thought, tool use, and multi-agent orchestration (ai.meta.com; About.fb). Coverage from Lifehacker and Built In notes Meta exposes multiple operating modes, described as Instant, Thinking, and Contemplating, to trade off latency against depth of reasoning. Meta also emphasizes a smaller, efficiency-focused architecture intended to enable low-latency responses on devices like glasses.

Industry pattern: compact, multimodal models that prioritize latency and on-device inference are a practical response to wearable constraints, where network round-trips and power budgets matter. Companies pursuing device-first experiences commonly reduce parameter count and add orchestration for tool access and multimodal reasoning to preserve capability while lowering compute and latency.

Coverage frames the model as Meta's effort to close gaps after the Llama series fell behind leading models from OpenAI, Google DeepMind, and Anthropic (UploadVR; Built In). Notably, Muse Spark is Meta's first proprietary, closed-weight model, a shift from its prior open-weights Llama strategy, though Meta has said it may consider open-sourcing future versions. For practitioners, the combination of multimodality, agentic tool use, and efficiency-focused design underscores a broader trade-off between open models and tightly integrated, product-first ones.

• •Real-world latency and power consumption on glasses hardware once rollouts reach users (About.fb).
• •Independent benchmark comparisons beyond the early leaderboard numbers cited by UploadVR.
• •Meta's future open-source commitments for the Muse family, since availability and licensing will shape developer adoption (UploadVR; ai.meta.com).

Meta described Muse Spark as the first model in a new Muse series and framed its plan as scaling deliberately, validating each generation before building larger ones (ai.meta.com; About.fb).