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 precocious AI probe division, connected Aug. 10 launched Muse Glimmer arsenic an open-weight agentic model designed to tally straight connected idiosyncratic devices, releasing its weights nether the permissive Apache 2.0 license. Open weights let developers to download a model’s underlying parameters, portion agentic refers to systems that implicit multi-step tasks utilizing tools with constricted quality input. The exemplary is distilled from Muse Spark, Meta’s larger closed model, and runs connected a Mac oregon PC with a azygous user graphics processing unit, oregon GPU.

The institution described:

“Muse Glimmer is simply a 30-billion-parameter exemplary optimized for always-on section cause workflows.”

Muse Glimmer completes tasks from commencement to finish, uses bundle tools, writes code, follows aggregate reasoning steps and recovers erstwhile an attempted enactment fails. The exemplary processes substance and images, handles information spanning much than 100 languages, and performed powerfully for its size people against Gemma4-31B and Qwen3.6-27B crossed agentic, coding, multimodal, information and reasoning benchmarks, according to Meta.

AI agents disagree from accepted chatbots done their quality to execute tasks, usage outer tools and marque decisions crossed extended workflows with varying levels of autonomy.

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

Running a exemplary of Muse Glimmer’s size connected user hardware required Meta to trim representation demands portion retaining show crossed tasks handled by AI agents. A 30-billion-parameter exemplary astatine afloat precision would transcend 55 GB, but 4-bit quantization, which stores exemplary information successful a much compact numerical format, shrank the connection exemplary beneath 20 GB and acceptable its components wrong a 24 GB oregon 32 GB envelope.

A smaller companion exemplary shipping alongside Muse Glimmer predicts upcoming pieces of text, letting the main exemplary verify respective predictions astatine erstwhile and respond faster. The technique, known arsenic speculative decoding, accrued procreation velocity by 3.1 times connected an RTX 5090, 1.8 times connected an M5 Max and 1.5 times connected an M4 Max successful institution testing.

The propulsion toward section agents arrives arsenic Meta expands spending and reorganizes teams astir autonomous AI products and faster improvement cycles. Capital expenditures for 2026 are expected to scope $130 cardinal to $145 billion, a scope narrowed from $125 cardinal to $145 cardinal erstwhile the institution reported second-quarter results connected July 29.

Roughly 7,000 employees moved into 4 AI-focused organizations successful May, including Applied AI Engineering and an Agent Transformation Accelerator team, portion astir 8,000 workers were laid disconnected and 6,000 unfastened roles closed. Meta aboriginal told transferred unit they could determine whether to stay.

Local Processing Reduces Dependence connected Continuous Cloud Access

Keeping cause workflows connected idiosyncratic hardware avoids routing each petition done distant information centers, allowing tasks to proceed erstwhile connectivity is constricted oregon unavailable. Local execution besides suits usage cases built connected stored discourse specified arsenic schedules, messages and files, wherever the exemplary acts connected idiosyncratic information crossed extended sessions.

Security standards are becoming progressively applicable arsenic autonomous bundle receives support to interact with accounts, applications and delicate information. The National Institute of Standards and Technology launched an AI Agent Standards Initiative covering unafraid connection betwixt agents, reliable integer identities, entree controls and communal method standards.

Meta emphasized:

“Running models locally enables you to usage AI anywhere, anytime, with oregon without an net connection.”

Beyond section models, Meta has moved into platforms built astir interactions among autarkic bundle agents. The institution acquired Moltbook successful March, an AI-agent societal network with much than 206,000 human-verified agents retired of astir 2.9 cardinal registered accounts arsenic 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 main executive, separately outlined a strategy connected Aug. 10 to distribute idiosyncratic superintelligence broadly and supply escaped versions to billions of people. Zuckerberg stated successful an Instagram video shared the aforesaid time that Meta would besides unfastened the weights for Muse Spark 1.2, framing the determination successful a 6,500-word effort arsenic an reply to open-weight releases from Chinese developers.

Supporting progressively susceptible models inactive requires extended centralized computing for grooming adjacent erstwhile the finished exemplary runs connected a user’s ain device. Meta signed a five-year statement with Nebius worthy up to $27 billion, including $12 cardinal successful dedicated capableness and options for different $15 billion.

Developers tin download Muse Glimmer’s weights now, portion optimized enactment for llama.cpp, MLX and ExecuTorch, bundle utilized to tally AI models efficiently crossed antithetic hardware, is expected successful the coming days. Meta is besides moving with AMD, Arm, Dell, Intel and Nvidia to optimize show crossed devices.