Meta 开源 Muse Glimmer：30B 本地代理模型，优化设备端执行
原标题：Meta Open-Sources Muse Glimmer: A 30B Local Agentic Model Optimised for On-Device Execution
AI 摘要
Meta AI Research 开源了 Muse Glimmer，一个 300 亿参数、Apache 2.0 许可的模型，专为本地设备上的自主代理、工具调用和编码任务设计。该模型通过动态量化和 DFlash 投机解码优化，可在消费级 GPU 上运行，内存占用降至 17-20 GB。Muse Glimmer 在多个基准测试中表现优于同类模型，并支持多种本地框架，标志着向高性能本地 AI 代理的转变。
正文节选
Meta AI Research has announced Muse Glimmer, a 30-billion-parameter open-weight model released under the Apache 2.0 license. Engineered specifically for always-on local workflows, Muse Glimmer enables developers to run autonomous agents, complex tool invocation, local coding, and LLM-as-a-judge evaluations directly on consumer GPUs and workstations without depending on cloud APIs. Muse Glimmer 30B model architecture and agentic benchmarks. Source: Sebastian Raschka To deliver agentic execution w