AI BriefingMeta Releases Muse Glimmer for On-Device AI
Key point
Meta has released the Muse Glimmer model and ExecuTorch support for on-device agentic workflows.
Details
Meta has released Muse Glimmer, a 30-billion-parameter (30B) open-weight model optimized for on-device agentic workflows. This model was distilled from Muse Spark and supports end-to-end execution on NVIDIA GPUs and Apple silicon-based Macs via ExecuTorch.
ExecuTorch adopts an approach of exporting models implemented in PyTorch as-is and optimizing them for each backend, rather than reimplementing models in other languages. This allows for optimization of the entire execution path using Triton in CUDA environments and MLX and Metal on Apple silicon.
Key Features and Capabilities:
- Multimodal Support: Capable of processing text and image inputs
- High-Performance Decoding: Low latency provided by applying DFlash (diffusion-based parallel speculative decoding)
- Flexible Deployment: Supports direct GGUF export and native K-quant execution
- Large Context: Supports token contexts of 128K or more
Validated PTE (PyTorch Executable) artifact bundles for NVIDIA CUDA and Apple silicon are currently available via Hugging Face, allowing users to download and run them immediately.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Meta Admits AI Assistant 'Muse' Was Heavily Inspired by OpenClaw