Meta CEO Mark Zuckerberg on Monday released a new open-weight model Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card. He also said that Meta supports open source AI and that he believes ‘everyone should have access to superintelligence.’
“Today we’re also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally,” Zuckerberg said on the social media platform X.
“I believe everyone should have access to superintelligence, and I wrote a long piece about Meta’s philosophy and values for building a positive future for everyone,” he added.
Muse Glimmer has been developed by Meta Superintelligence Labs and is specifically optimised for agentic AI tasks. Unlike a traditional chatbot that mainly responds to questions, an AI agent can plan tasks, use external tools, check its own work and recover when something goes wrong.
According to Meta, Muse Glimmer can handle tasks such as coding, function calling, local file management and other multi-step workflows. It can also work with images through a dedicated perception encoder, allowing it to understand screenshots, charts and documents along with text.
“Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows. It’s small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category,” the company said in a blog post.
According to the blog post , the Social media giant has used several techniques to make the model small enough for local deployment. The tech giant has compressed its weights to around 4-bit precision, bringing the language model to under 20GB.
As per the company, Muse Glimmer was trained using outputs from Meta’s larger Muse Spark model, followed by additional training focused on longer contexts, reasoning, coding and agentic tasks.
Meta says the model has been evaluated across benchmarks covering tool use, coding, reasoning and end-to-end agent performance.
The model supports more than 100 languages and is designed to work with popular AI development tools and frameworks. The weights are available on Hugging Face, while integrations with platforms including Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang are expected.
Meta is also working with companies including AMD, Arm, Dell, Intel and Nvidia to improve Muse Glimmer’s performance across different devices.
Alexandr Wang, Chief AI Officer at Meta, also shared the model’s ability to run on consumer hardware. “Muse Glimmer can run on 24GB of VRAM without losing agentic reliability,” Wang wrote on X.