Meta launches Muse Glimmer to run AI agents locally

Meta released Muse Glimmer on Aug. 10, a 30‑billion‑parameter, open‑weight AI agent designed to run locally on a Mac or PC with a single consumer GPU. The model’s weights were published under an Apache 2.0 license by Meta Superintelligence Labs and are available for developers to download now.

Muse Glimmer is a distilled version of Meta’s larger Muse Spark model. It is built to perform multi‑step tasks end to end: interacting with software tools, writing code, following chains of reasoning and recovering when an attempted action fails. The model accepts text and images, supports more than 100 languages and, according to Meta, matched or exceeded similarly sized models on agentic, coding, multimodal, safety and reasoning benchmarks.

Meta reduced Muse Glimmer’s memory needs using 4‑bit quantization, shrinking a model that would exceed 55 GB at full precision to under 20 GB. That allows the model and its components to run on systems with 24 GB or 32 GB of GPU memory. A smaller companion model ships with Muse Glimmer to predict upcoming tokens so the main model can verify multiple predictions at once. Meta reports speculative decoding increased generation speed by 3.1× on an RTX 5090, 1.8× on an M5 Max and 1.5× on an M4 Max in company tests.

Running the model locally keeps processing on users’ devices rather than sending every request to cloud servers. Meta highlighted: “Running models locally enables you to use AI anywhere, anytime, with or without an internet connection.” Local execution supports tasks that rely on stored context such as schedules, messages and files and can continue when network access is limited or unavailable.

Local agents that access accounts and sensitive data raise security and identity questions. The National Institute of Standards and Technology has launched an AI Agent Standards Initiative to address secure communication between agents, reliable digital identities, access controls and common technical standards.

Developers can download Muse Glimmer’s weights immediately. Meta said optimized support for runtimes such as llama.cpp, MLX and ExecuTorch will arrive in the coming days. The company is working with chipmakers and system vendors including AMD, Arm, Dell, Intel and Nvidia to improve performance across devices.

The release comes as Meta adjusts its organization and capital plans around AI. The company expects capital expenditures for 2026 to fall between $130 billion and $145 billion. In May, Meta moved about 7,000 employees into four AI‑focused organizations, while roughly 8,000 positions were eliminated and 6,000 open roles closed. Meta signed a five‑year agreement with cloud provider Nebius worth up to $27 billion to secure training capacity. The company acquired Moltbook, an AI‑agent social network, earlier this year and indicated plans to make more advanced models broadly available; Meta’s CEO announced that weights for Muse Spark 1.2 will be opened.