Meta has announced the release of Muse Glimmer, a new open-source artificial intelligence model engineered to run efficiently on personal computers. This lightweight model is a distilled version of Meta's larger Muse Spark 1.2, featuring 30 billion parameters, and is optimized for local use cases, allowing users to access AI functionalities with or without an internet connection.
Agentic Capabilities for Local Tasks
Unlike traditional AI models that primarily provide responses, Muse Glimmer is designed to perform agentic-like tasks directly on a user's behalf. Its capabilities include managing schedules, organizing files, and handling other multi-step computer tasks locally. Meta emphasizes that this optimization enables users to leverage AI anywhere, anytime, ensuring privacy and accessibility.
Technical Specifications and Accessibility
Muse Glimmer is released under a permissive Apache 2.0 license, making it broadly accessible for developers and researchers. While the model initially required over 55GB of memory, Meta utilized quantization techniques to compress its weights to approximately 4-bit precision, reducing its size to under 20GB. This compression allows Muse Glimmer to run efficiently on a single consumer-grade Graphics Processing Unit (GPU) found in many modern PCs and Macs.
The model was trained through a multi-stage process, primarily on outputs from the Muse Spark model using logit distillation, followed by mid-training on agent-heavy data and post-training with supervised fine-tuning and reinforcement learning across reasoning, coding, and agentic domains.
Performance and Integration
Meta reports that Muse Glimmer demonstrates strong performance across various benchmarks, including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which evaluate coding, tool use, and multi-turn task completion. The model can perform multi-step reasoning, make and recover from tool calls, and understands both text and images, having been trained on data spanning over 100 languages.
Muse Glimmer is readily available via Hugging Face and supports local deployment platforms such as Ollama, LM Studio, and Unsloth. Frameworks like llama.cpp, ExecuTorch, and MLX are also set to support the model. Meta is actively collaborating with hardware companies including AMD, Arm, Dell, Intel, and Nvidia to further optimize Muse Glimmer across a diverse range of devices, enhancing its reach and performance on various systems.