Meta's Muse Glimmer Puts an Open AI Agent on Your Laptop

Meta released Muse Glimmer on Monday, an open-weight artificial intelligence model small enough to run on a single consumer graphics card while still handling the kind of multi-step "agent" work that has mostly required a data centre until now. The company published the model on Hugging Face under the permissive Apache 2.0 licence, so developers can download the weights, inspect them and build on them without paying Meta or asking permission.

Muse Glimmer has 30 billion parameters and is a distilled version of Meta's larger Muse Spark 1.2 model. At full precision it would need more than 55GB of memory, but 4-bit compression shrinks it below 20GB, enough to run on a well-equipped laptop or a desktop with a 24 to 32GB card such as an RTX 5090. Tools that host models locally, including Ollama, LM Studio and vLLM, supported it from the first day.

Built for local agents

Where earlier small models were pitched mainly at chat, Meta is aiming Muse Glimmer at agentic tasks: calling tools, working through a problem in several steps, recovering from its own mistakes, and taking images as well as text. It also lets developers dial the amount of "reasoning effort" up or down depending on the job. Meta describes the target user as someone building a local coding assistant, a function-calling system, or an automated evaluator that grades other models, all without a constant connection to the cloud.

That puts Muse Glimmer squarely against Google's Gemma and Alibaba's Qwen family, the two open lines that have set the pace for efficient, self-hosted AI this year. It also lands the same week that ByteDance confirmed it is training a 10-trillion-parameter model, a reminder that the industry is now pushing hard at both ends of the scale at once.

A launch with a lobbying message

Mark Zuckerberg used the release to make a policy argument in Washington. His case is that American labs are at a structural disadvantage to foreign rivals, particularly over restrictions on the data they can train on, and he called for those compliance burdens to be eased. He rejected proposals to bar foreign open-source models from the United States, and defended distillation, the technique Meta used to compress Muse Spark into Glimmer, as a legitimate and valuable practice.

The stance separates Meta from closed-model advocates such as OpenAI and Anthropic, though the company has said it still expects to hold back its most capable "superintelligence" systems rather than release them openly. The question of how freely powerful models should circulate is already live in Washington, where regulators recently weighed curbs on Chinese open-weight models even as they eased the path for American ones.

For developers, the immediate significance is simpler. A capable, tool-using model that runs on hardware they already own lowers both the cost and the privacy risk of building AI into everyday software. Whether Glimmer holds up against the cloud-hosted giants in real use is the question the next few weeks will answer.