Meta launches Muse Glimmer to bring AI agents closer to devices
Meta has launched Muse Glimmer, a 30-billion-parameter open model designed to run AI agents locally on consumer hardware.
AI agents are moving beyond the cloud. Meta has introduced Muse Glimmer, a 30-billion-parameter open agentic model designed to run on consumer hardware.
Announced by Meta Superintelligence Labs on 10 August 2026, the model is being released with open weights under the Apache 2.0 licence, allowing developers to study, adapt and build applications around it.
Why local AI agents matter
Muse Glimmer is designed for agentic AI, where models can plan tasks, use software tools and work through multiple steps to reach a goal.
Unlike a conventional chatbot that mainly responds to individual prompts, an AI agent can decide which tool to use, carry out an action, check the result and continue working. Running that process locally could offer several advantages.
A model operating directly on a Mac or PC does not always need to send data to a remote server, potentially improving privacy and reducing dependence on an internet connection.
Built for longer, more complex tasks
Meta says Muse Glimmer has been optimised for local agents, coding, function calling, and LLM-as-a-judge evaluation. Function calling allows an AI model to interact with software tools, while LLM-as-a-judge uses another language model to assess the quality of an AI-generated response.
The model is also trained for long-horizon execution, multimodal understanding, long-context memory and instruction following. In practical terms, that means it is designed to keep track of longer tasks, work with text and images, interact with tools and maintain the user's instructions across several steps.
How Meta trained the model
Meta says Muse Glimmer was developed through three main training stages. The first used logit distillation from Muse Spark. In this process, a smaller model learns from the output patterns of a larger model.
In the second stage, training expanded to longer-context and more agent-focused data, including detailed reasoning traces. The final stage combined supervised on-policy distillation, fine-tuning, and reinforcement learning across coding, reasoning, and agentic tasks.
Together, these steps were designed to create a model capable of useful agent workflows while keeping its hardware requirements within reach of consumers.
Fitting 30 billion parameters on a PC
Memory is one of the biggest challenges when running large AI models locally. Meta says a full-precision version of Muse Glimmer would require more than 55 GB of memory. The company uses quantisation, which compresses model weights, to bring the language model below 20 GB.
That leaves space within a 24 GB or 32 GB hardware setup for working memory, image-processing components and a speculative decoding model. Speculative decoding can improve response speed by allowing a smaller model to suggest groups of tokens that the main model can quickly check.
Meta says this approach can make Muse Glimmer substantially faster than conventional token-by-token generation while maintaining output quality.
Developers can start experimenting
Muse Glimmer is available through Hugging Face, with Meta also providing documentation for developers.
The company says support is planned for tools and frameworks including llama.cpp, MLX, ExecuTorch, etc.
It also lists partners such as Ollama, LM Studio, Unsloth, Together AI, Fireworks AI and OpenRouter. For developers, the model provides another option for experimenting with AI agents that can run closer to the user.
For Meta, it strengthens the company's push towards open-weight AI that is not limited to large cloud platforms. If local hardware continues to become more capable, models such as Muse Glimmer could make agentic AI more practical on everyday computers.