Meta AI Releases Open-Weights Agentic Model Muse Glimmer

Meta AI released Muse Glimmer, a 30-billion-parameter multimodal model, on an unspecified recent date, making it available under the Apache 2.0 license. This model is a distillation from Muse Spark and is specifically tuned for always-on local agent workflows, enabling applications to run entirely on user hardware without network calls. Normally, a 30-billion-parameter model requires over 55 GB of memory at full precision. Meta has compressed Muse Glimmer to approximately 4-bit precision and incorporated block-level speculative decoding to ensure fast response times, making it suitable for integration into real-time agent loops. The model's design allows it to operate on a single consumer GPU or a Mac computer, eliminating the need for cloud-based processing. This open-weights release facilitates self-hosting, with weights available in BF16 format, GGUF k-quants, ExecuTorch builds, and a DFlash drafter on Hugging Face. Solo developers and startups can deploy Muse Glimmer on a 24 GB GPU or an M4/M5 Max Mac. Mid-market teams can achieve on-premise inference without incurring per-token costs, while regulated enterprises can implement an air-gappable agent solution. Meta recommends implementing system-level guardrails for enhanced security rather than exposing the model as a raw endpoint. The model is particularly relevant for industries such as healthcare, legal services, financial services, defense, the public sector, manufacturing, and field service, where data residency requirements, offline operation needs, or strict latency constraints preclude cloud-based solutions. Potential applications include desktop agents capable of reading screenshots, coding assistants, and schema-based function calling. It can also be used for document and chart understanding, generating synthetic data, and performing LLM-as-a-judge evaluations. Muse Glimmer is architected as a dense causal transformer, featuring a dedicated perception encoder. Its total parameter count of approximately 30 billion includes a vision tower. The model employs grouped-query attention with 32 query heads and 2 KV heads. Its attention mechanism follows a repeating pattern of [Local, Local, Local, Global] with a 2,048 token sliding window. Rotary Position Embeddings (RoPE) are applied exclusively to local layers, utilizing a theta value of 500,000. The vision component of the model is a ViT-G/14 percept, comprising roughly 1.8 billion parameters.