Meta Releases Muse Glimmer, 30B On-Device Agentic Model via ExecuTorch

Meta introduced Muse Glimmer, an open-weight, 30-billion-parameter model distilled from its Muse Spark for on-device agentic workflows¹. ExecuTorch now provides end-to-end support for running Muse Glimmer on NVIDIA GPUs and Macs with Apple silicon, with prebuilt artifacts published on Hugging Face for both platforms. The model supports text and image inputs, 128K+-token context, native K-quant execution, direct GGUF export, and DFlash parallel diffusion-based speculative decoding. Its KV-cache uses 13 global and 39 sliding-window layers across 52 total layers for efficient memory scaling.