Meta launches Muse Glimmer, a 30B open-weight AI model designed to run locally

Meta has introduced Muse Glimmer, a 30-billion-parameter artificial intelligence model designed to run agentic workloads directly on consumer devices, as the company steps up its push for open-weight AI models.

Developed by Meta Superintelligence Labs, Muse Glimmer is being released with its model weights under the permissive Apache 2.0 licence. Meta said the model is small enough to run on a Mac or PC equipped with a single consumer graphics processing unit (GPU), targeting applications including local AI agents, function calling, coding and AI-based evaluation. 

The announcement comes as Meta seeks to position locally run AI as an alternative to models that rely heavily on cloud infrastructure and continuous internet connectivity. Running AI models on-device can allow users to access AI capabilities even when they are offline, while potentially reducing the need to send sensitive information to external servers.

“Running models locally enables you to use AI anywhere, anytime, with or without an internet connection,” Meta said in its announcement. 

Muse Glimmer is specifically built for agentic workflows, where an AI model does more than respond to prompts and can plan and execute a sequence of actions to complete a task. Meta said the model can handle long-horizon execution, tool calling, multimodal inputs, long-context memory and instruction following.

What can Muse Glimmer do?

The model can perform end-to-end tasks, use external tools, sustain multi-step reasoning and recover from failed tool calls. It can also process text and images, allowing an agent built on Muse Glimmer to interpret screenshots, charts and documents. The model has been trained on data spanning more than 100 languages. 

Meta said Muse Glimmer was trained using outputs from its larger Muse Spark model through a process called logit distillation. The company subsequently trained it on longer-context and agent-focused data, before using supervised fine-tuning, additional distillation and reinforcement learning across reasoning, coding and agentic tasks.

A key challenge was getting the 30-billion-parameter model to operate within the memory constraints of consumer hardware. At full precision, Meta said the model would require more than 55 GB of memory. It therefore used quantisation to bring the model down to under 20 GB at approximately 4-bit precision, allowing it to operate within a 24 GB or 32 GB memory envelope alongside other components required for image processing and inference.

Meta has also used speculative decoding to improve the model’s response speed. The company said its tests showed decode-speed improvements of 3.1 times on an Nvidia RTX 5090, 1.8 times on Apple’s M5 Max and 1.5 times on the M4 Max when paired with its DFlash-based drafter. 

The model is available through Hugging Face, while integrations with tools including llama.cpp, MLX and ExecuTorch are expected in the coming days. Meta also said developers will be able to run Muse Glimmer through platforms including Ollama, LM Studio and Unsloth, as well as deploy it using frameworks such as vLLM and SGLang.

Meta is working with chip and hardware companies including AMD, Arm, Dell, Intel and Nvidia to optimise the model across different devices.

The announcement comes alongside Meta’s broader effort to reassert itself in the AI race. CEO Mark Zuckerberg has recently reaffirmed the company’s support for open-weight AI, arguing that more accessible models could help developers compete with closed systems from companies such as OpenAI and Anthropic.