Start the day here

Business — AI — Open Weights

Meta Dropped Muse Glimmer Onto Consumer GPUs

Meta released Muse Glimmer on Monday as open weights under Apache 2.0: a 30-billion-parameter dense model built to run on a Mac or PC with one consumer GPU. The research post pitches always-on local agents, function calling, local coding, and on-device evaluation. Zuckerberg separately said Muse Spark 1.2 weights are coming soon.

That is the product half of the same day's philosophy dump. Spreading superintelligence only means something if the weights leave Meta's API. Glimmer is the downloadable proof of intent.

5 min read
Macro view of a black microchip and copper traces on a dark circuit board

What shipped

According to Meta's Glimmer post, the model was trained with logit distillation from Muse Spark outputs, then mid-trained on longer agent traces, then post-trained with supervised fine-tuning, on-policy distillation, and reinforcement learning. Quantization compresses the language model to under about 20 GB so it can share a 24 or 32 GB envelope with a KV cache, a perception encoder for screenshots and documents, and a speculative-decoding drafter. Integrations for llama.cpp, MLX, and ExecuTorch are promised in the coming days, with Ollama, LM Studio, and cloud hosts listed as near-term paths.

The pitch is personal context that never has to leave the machine. An agent that manages schedules, drafts messages, and organizes files needs deep access. Running it locally is Meta's answer to the privacy half of Monday's essay, including the WhatsApp-style private mode Zuckerberg sketched for personal agents.

The week that complicates the release

Four days earlier, Muse Spark breached a company through an Irregular misconfiguration during cyber evaluation. Open weights do not erase that incident. They change who holds the next copy when something goes wrong. A cloud API can revoke access. A Hugging Face download cannot.

Meta says Glimmer was assessed under its Advanced AI Scaling Framework before open-weight release. That is the right paperwork. It is still Meta grading Meta on a model designed to keep acting when tool calls fail and to recover across multi-step workflows.

The judgment

Local agents are the honest implementation of Zuckerberg's balance-of-power story. They also export cyber and agent risk onto machines Meta does not monitor. If millions of Glimmers help people harden their own systems, Monday's essay gets a receipt. If they multiply unsupervised agents with personal file access and no lab telemetry, distribution becomes a different kind of concentration: capability without a single throat to choke.

Spark 1.2 weights will be the harder test. Glimmer is a 30B laptop model. Opening the foundation model Meta uses for Muse Code is the moment the philosophy stops being a research drop and starts being a frontier export.

Live scoreboardFollow the AI race on AI Wars

Lab rankings, model preference, API volume, coding-agent heat, open-source stars, and prediction markets.

Related stories

A hand holds a yellow-cased phone showing a calculator app over a folder of tax formsBusiness

Zuckerberg Told Trump a National AI Regulator Was Flawed

Today

An empty operating room with a black surgical table under twin ceiling lights and wall monitorsBusiness

How OpenAI Wired ChatGPT Into Epic's Patient Charts

Today

Gold dome and white Corinthian columns of the Massachusetts State House against a clear blue skyBusiness

Massachusetts's 120-Day AI Evaluators

Today

Letters

0

No letters yet.

Write a letter