Meta Deploys Muse Spark On Smart Glasses

According to Meta's product blog, the company has started rolling out Muse Spark across the Meta AI experience and is deploying it on smart glasses hardware, with the rollout of Muse Spark on Ray-Ban Meta and Oakley Meta in the US and Canada and Meta Ray-Ban Display slated for summer (About.fb; ai.meta.com). UploadVR and Built In report that Muse Spark is the first model released by Meta Superintelligence Labs, and those outlets 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 60-61 for newer competitor models.

Per Meta's public materials, Muse Spark is described as a natively multimodal reasoning model 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 that Meta exposes multiple operating modes in the model, described as Instant, Thinking, and Contemplating, to trade off latency and depth of reasoning in client experiences (Lifehacker; Built In). Meta's communications also emphasize a smaller, efficiency-focused architecture intended to enable low-latency responses on devices like glasses (About.fb; UploadVR).

Industry-pattern observations: 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 similar device-first experiences commonly reduce model parameter count and add orchestration for tool access and multimodal reasoning to maintain capability while lowering compute and latency requirements.

Public coverage frames this update as Meta's effort to close capability gaps created by the Llama series falling behind leading models from OpenAI, Google DeepMind, and Anthropic, a narrative captured in UploadVR and Built In reporting. The decision to keep this initial Muse Spark release closed-source while stating an intent to consider open-sourcing future versions is flagged in multiple outlets as a notable change from how Meta handled prior Llama releases (UploadVR; Built In). For practitioners, the combination of multimodality, agentic tool use, and efficiency-focused design underscores a broader trade-off trend between open models and tightly integrated, product-first models.

For practitioners and product teams, observers will likely monitor three indicators: real-world latency and power consumption on glasses hardware once the rollouts reach users (About.fb); independent benchmark comparisons beyond the early leaderboard numbers cited by UploadVR; and Meta's future open-source commitments for the Muse family, since availability and licensing will shape developer adoption and replication studies (UploadVR; ai.meta.com).

Meta Superintelligence Labs wrote in its public announcement, "Muse Spark is the first model in our new Muse series, a deliberate and scientific approach to model scaling where each generation validates and builds on the last before we go bigger" (ai.meta.com; About.fb).