Meta Open-Sources New 30B Model to Run AI Agents Locally on Devices

Meta has introduced Muse Glimmer, a 30-billion-parameter model to power always-on AI agents directly on consumer devices. Developed by Meta Superintelligence Labs, Muse Glimmer is designed to run locally on a Mac or PC equipped with a single consumer GPU.

It is working with AMD, Arm, Dell, Intel and NVIDIA to optimise the model across different hardware platforms.

"Foundation models have achieved remarkable capabilities across reasoning, code generation, and tool use — yet most deployments still depend on cloud infrastructure and network access," Meta said in a blog post.

Meta is positioning the model as a smaller alternative to cloud-dependent AI systems, targeting applications including local agents, coding, function calling and automated model evaluation.

The company said Muse Glimmer was built around the idea that AI agents should be capable of operating without continuous access to cloud infrastructure or an internet connection.

Local deployment could also give users greater control over personal information, particularly for agents handling tasks such as managing schedules, organizing files and drafting messages.

"Running models locally enables you to use AI anywhere, anytime, with or without an internet connection. This is increasingly viable: the open source community has shown that smaller models, when trained effectively, can approach frontier-level performance on targeted tasks. Muse Glimmer is optimised for these local use cases," Meta added.

Meta trained Muse Glimmer using outputs from its larger Muse Spark model through a distillation process. The training involved multiple stages covering long-context and agent-focused data, supervised fine-tuning, reinforcement learning, coding, reasoning and tool use.

The model supports multimodal inputs, allowing agents to process text alongside images such as screenshots, charts and documents. It is also trained across more than 100 languages and includes capabilities for multi-step reasoning, tool calling and recovering from failed tool calls.

A key challenge was fitting the 30-billion-parameter model onto consumer hardware. Meta said quantization reduces the model to under 20GB, allowing it to operate within a 24GB or 32GB memory envelope. The company has also integrated speculative decoding to accelerate generation and make agent interactions more responsive.

Muse Glimmer is available through Hugging Face, with integrations for llama.cpp, MLX and ExecuTorch expected soon. Developers will also be able to run it through platforms including Ollama, LM Studio and Unsloth, while deployment options include vLLM and SGLang.

Recently, Meta announced that its latest AI reasoning models achieved top-tier results across five of the world's most prestigious STEM Olympiad competitions.