Meta has released the weights of Muse Glimmer, a 30-billion-parameter artificial intelligence (AI) model designed to run locally on personal computers. According to the US-based technology company, the model is built for AI agents, coding and tool use, and can run on a Mac or PC with a single consumer graphics processing unit (GPU).
The release comes as AI models are increasingly being developed for local use, allowing workloads to run on devices rather than depending entirely on cloud infrastructure.
Meta said Muse Glimmer is optimised for local agent workflows and can handle tasks such as function calling, multistep reasoning and coding. The company has released the model's weights under the Apache 2.0 licence.
Mark Zuckerberg also confirmed the release in a post on X.
"Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally," he said.
Zuckerberg added that Meta would soon release the weights of Muse Spark 1.2, which he described as the company's latest foundation model.
What is Meta Muse Glimmer
Muse Glimmer is a 30-billion-parameter AI model designed to run locally on a Mac or PC with a single consumer GPU. Meta said it has optimised the model for local AI agents that can work continuously on tasks without depending entirely on cloud infrastructure or an internet connection.
The model can be used for local AI agents, function calling, coding and large language model (LLM)-as-a-judge evaluation.
Meta said Muse Glimmer performs strongly on key agentic tasks and benchmarks compared with other models in the same size category.
The company evaluated Muse Glimmer across a range of benchmarks covering capabilities required by autonomous AI agents. Meta said the model performed strongly for its size class on several widely used LLM benchmarks compared with Gemma 4 31B and Qwen 3.6 27B.
Why Meta is focusing on local AI
Most AI models are still commonly accessed through cloud-based services, which require an internet connection. Running an AI model locally allows users to access it even when they are offline while reducing dependence on remote cloud infrastructure.
Meta said advances in training smaller models have made local AI increasingly practical. According to the company, smaller models can deliver strong results on specific tasks when trained effectively.
Muse Glimmer has been developed with these local use cases in mind. The model is also designed to work with tools developers already use. Meta said integrations with llama.cpp, MLX and ExecuTorch will be available in the coming days.
How Meta trained Muse Glimmer
Meta said developing a model for local use required balancing performance with the memory and computing constraints of personal computers.
The company used a compact architecture, distillation and inference optimisations, including quantisation.
Training was carried out in three stages.
- During pre-training, Meta used outputs from its Muse Spark model through logit distillation, with a data mix similar to that used for the larger model.
- During mid-training, Muse Glimmer was trained on longer-context and more agent-focused data, including richer reasoning traces and organic data.
- For post-training, Meta combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding and agentic tasks.
Meta said the training was designed to provide Muse Glimmer with capabilities required for local AI agents, including long-horizon task execution, tool use, multimodal understanding, long-context memory and instruction following.
The company evaluated Muse Glimmer under its Advanced AI Scaling Framework before releasing the model with open weights.
What Meta Muse Glimmer can do
Complete tasks end to end: Muse Glimmer is designed to handle complete tasks, including searching for information, writing and debugging code, and resolving requests across multiple turns.
Use tools: The model can make different types of function calls and use tools with specific schemas during longer workflows.
Handle multistep reasoning: Muse Glimmer can maintain reasoning and follow plans across complex tasks requiring several steps.
Recover from errors: If a tool call fails or returns an unexpected result, the model can identify the error and retry instead of stopping.
Process text and images: A dedicated perception encoder allows the model to work with text and images, including screenshots, charts and documents.
Work with agent frameworks: Muse Glimmer is compatible with OpenClaw and other agentic orchestration frameworks.
Adjust reasoning effort: The model supports different levels of reasoning, allowing users to balance response quality with speed.
Support more than 100 languages: Muse Glimmer has been trained on data covering more than 100 languages.
Meta Muse Glimmer: Availability
Muse Glimmer is available now, with its weights available for download on Hugging Face.
Meta said developers will be able to run the model locally through Ollama, LM Studio and Unsloth in the coming days.
The model can also be deployed using edge frameworks such as llama.cpp, ExecuTorch and MLX, while vLLM and SGLang can be used to serve it at scale.
Developers can also access Muse Glimmer through Together AI, Fireworks AI and OpenRouter.
Meta said developers can customise the model for specific use cases using PyTorch's TorchTitan training framework.
The company is also working with AMD, Arm, Dell, Intel and Nvidia to optimise the model's performance across devices.
Meta has released documentation to help developers set up and build with Muse Glimmer, including guidance on creating custom scaffolds for personal AI agents. The company said additional resources are available through its AI Developer Center.
 "Muse Glimmer 30B: Meta bets on open AI models that can run on Macs and PCs")
