Meta moves multi‑agent Muse Spark architecture onto consumer hardware

Meta said it is shipping Muse Spark 1.3 (Contemplating mode) to consumer devices, running up to 16 parallel reasoning agents to handle complex, low‑latency tasks on glasses and VR hardware.

In this brief: 2 sections 1 min read

- Contemplating mode can orchestrate up to 16 parallel reasoning agents to generate, refine and aggregate candidate solutions.
- Meta said Muse Spark 1.3 uses multi‑round test‑time scaling to manage complex reasoning without proportional latency increases.
- The architecture was moved from API/datacenter deployments onto Ray‑Ban Meta updates and Project Phoenix hardware at Connect.
- Meta continues to train large models on datacenter accelerators (Iris/MTIA program), not consumer chips.
- Consumer devices (Project Phoenix, Ray‑Ban updates) rely on Qualcomm Snapdragon Reality Elite or similar edge compute.
- Model and architecture choices aim to bridge cloud training with edge inference for agentic use cases.

Read full analysis on forkast.news

Useful?