Meta发布轻量级AI模型Muse Ember,推理成本仅为Anthropic Claude Haiku 5.5的三分之一。
Meta has launched Muse Ember, a lightweight AI model explicitly optimized for edge devices and high-frequency interactions; its official inference cost is one-third that of Anthropic’s Claude Haiku 5.5. Already deployed on Apple Watch and Wear OS, it delivers low-latency voice and text responses with local context caching to minimize cloud round trips. Benchmarks show comparable accuracy to Haiku 5.5 on common agent tasks (e.g., calendar parsing, message summarization, multi-step command execution), yet with 40% lower P99 latency and 55% less memory usage. For early-stage AI application teams, this signals a shift toward ‘cost-aware agent deployment’: developers no longer need to choose crudely between small-model efficiency and large-model capability. Instead, they should adopt layered routing (e.g., simple queries → Ember, complex reasoning → Claude Opus), implement state-aware local caching, and treat cost metrics ($/1k tokens, cold-start time) as first-class MVP acceptance criteria — alongside accuracy and latency.