Meta Launches Muse Glimmer for Local AI Agents on Personal Devices

Muse Glimmer Brings AI Agents Directly Onto Consumer Hardware

Meta Superintelligence Labs, the company’s advanced AI research division, on Aug. 10 launched Muse Glimmer as an open-weight agentic model designed to run directly on personal devices, releasing its weights under the permissive Apache 2.0 license. Open weights allow developers to download a model’s underlying parameters, while agentic refers to systems that complete multi-step tasks using tools with limited human input. The model is distilled from Muse Spark, Meta’s larger closed model, and runs on a Mac or PC with a single consumer graphics processing unit, or GPU.

The company described:

“Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows.”

Muse Glimmer completes tasks from start to finish, uses software tools, writes code, follows multiple reasoning steps and recovers when an attempted action fails. The model processes text and images, handles data spanning more than 100 languages, and performed strongly for its size class against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning benchmarks, according to Meta.

AI agents differ from conventional chatbots through their ability to execute tasks, use external tools and make decisions across extended workflows with varying levels of autonomy.

Meta Compresses a 30-Billion-Parameter Model Below 20 GB

Running a model of Muse Glimmer’s size on consumer hardware required Meta to reduce memory demands while retaining performance across tasks handled by AI agents. A 30-billion-parameter model at full precision would exceed 55 GB, but 4-bit quantization, which stores model data in a more compact numerical format, shrank the language model below 20 GB and fit its components within a 24 GB or 32 GB envelope.

A smaller companion model shipping alongside Muse Glimmer predicts upcoming pieces of text, letting the main model verify several predictions at once and respond faster. The technique, known as speculative decoding, increased generation speed by 3.1 times on an RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max in company testing.

The push toward local agents arrives as Meta expands spending and reorganizes teams around autonomous AI products and faster development cycles. Capital expenditures for 2026 are expected to reach $130 billion to $145 billion, a range narrowed from $125 billion to $145 billion when the company reported second-quarter results on July 29.

Roughly 7,000 employees moved into four AI-focused organizations in May, including Applied AI Engineering and an Agent Transformation Accelerator team, while about 8,000 workers were laid off and 6,000 open roles closed. Meta later told transferred staff they could decide whether to stay.

Local Processing Reduces Dependence on Continuous Cloud Access

Keeping agent workflows on personal hardware avoids routing every request through remote data centers, allowing tasks to continue when connectivity is limited or unavailable. Local execution also suits use cases built on stored context such as schedules, messages and files, where the model acts on personal data across extended sessions.

Security standards are becoming increasingly relevant as autonomous software receives permission to interact with accounts, applications and sensitive information. The National Institute of Standards and Technology launched an AI Agent Standards Initiative covering secure communication between agents, reliable digital identities, access controls and common technical standards.

Meta emphasized:

“Running models locally enables you to use AI anywhere, anytime, with or without an internet connection.”

Beyond local models, Meta has moved into platforms built around interactions among independent software agents. The company acquired Moltbook in March, an AI-agent social network with more than 206,000 human-verified agents out of nearly 2.9 million registered accounts as of June, and brought its co-founders into Meta Superintelligence Labs.

Muse Glimmer Fits Meta’s Broader Push for Widely Available AI

Mark Zuckerberg, Meta’s chief executive, separately outlined a strategy on Aug. 10 to distribute personal superintelligence broadly and provide free versions to billions of people. Zuckerberg stated in an Instagram video shared the same day that Meta would also open the weights for Muse Spark 1.2, framing the move in a 6,500-word essay as an answer to open-weight releases from Chinese developers.

Supporting increasingly capable models still requires extensive centralized computing for training even when the finished model runs on a user’s own device. Meta signed a five-year agreement with Nebius worth up to $27 billion, including $12 billion in dedicated capacity and options for another $15 billion.

Developers can download Muse Glimmer’s weights now, while optimized support for llama.cpp, MLX and ExecuTorch, software used to run AI models efficiently across different hardware, is expected in the coming days. Meta is also working with AMD, Arm, Dell, Intel and Nvidia to optimize performance across devices.