EVENT DOSSIER
On August 10, 2026, Meta officially released the Muse Glimmer open-source model, which supports local multi-modal agent operations.
On August 10, 2026, Meta officially launched the Muse Glimmer model. This model features 30 billion parameters and a context window of over 120,000 words, designed for local operation and supporting multi-modal input and output of text, images, and audio. Muse Glimmer has autonomous planning capabilities and can run on NVIDIA’s edge devices, desktop, and workstation AI platforms. Its single-GPU inference speed reaches 20,000 tokens per second, enabling it to process local data and perform complex tasks in real time. The model is completely open-source, and its source code is available for community use.
Meta Muse Glimmer 发布开源本地多模态智能体模型
On August 10, 2026, Meta officially released the Muse Glimmer open-source model, which supports local multi-modal agent operations.
Both the NVIDIA Developer Blog and the Hugging Face Blog reported that Meta launched the Muse Glimmer model. This model has 30 billion parameters and over 120,000 context windows, and supports local multi-modal agent operations.
2 reportsMeta has launched Muse Glimmer, a localized, proxy-based, multi-modal, and open-source model. This model can run locally, has autonomous planning capabilities, can handle various input and output formats such as text, images, and audio, and its source code is fully open for community use.
Meta 发布 Muse Glimmer，这是一款拥有 300 亿参数、12 万 + 上下文窗口且专为本地 AI 代理工作设计的开源稠密模型。该模型优化后可在 NVIDIA 边缘、桌面及工作站 AI 平台上运行，单 GPU 推理速度达每秒 2 万 tokens，支持代理始终在线处理本地数据并执行复杂任务。