Meta’s Muse Glimmer fits on a laptop

Meta unveiled Muse Glimmer on Monday, a compact open-weight model optimized for local hardware to handle agentic workflows. Available on Hugging Face, Muse Glimmer emerges from Meta's compression of its larger Muse Spark model, allowing it to operate as a local agent. This approach demonstrates how companies can transform effective cloud models into smaller agents for local deployment, a process that Meta is integrating into a comprehensive pipeline.

While Glimmer can manage routine tasks on the device, a larger cloud model handles training and complex jobs, adding complexity for developers. Distillation, a technique that reduces model size, is used to convert Muse Spark into Glimmer while training it on Spark outputs using logit distillation, longer-context training focused on agents and reasoning, and post-training with supervised fine-tuning and reinforcement learning.

The resulting model boasts a 131,000-token context window, text and image input support, function calling, failure recovery, and multiple reasoning settings. However, running Glimmer on a typical laptop may require significant hardware costs due to its high memory demand. Meta provides quantized versions to address this issue, with the smallest configuration requiring 17GB of memory.

The model includes a "drafter" that predicts and verifies token blocks, improving generation speed on various hardware. Despite Meta's benchmarks placing Glimmer above its competitors, security remains the developer's responsibility. Developers must implement additional safeguards to prevent potential information mishandling within the local agent.