Meta releases 30B Muse Glimmer model for local AI agents on Macs and PCs

Meta has released Muse Glimmer, a 30-billion-parameter AI model built to run AI agents directly on supported Macs and PCs, including setups equipped with a single consumer GPU. Developed by Meta Superintelligence Labs, the model targets local workflows that would otherwise depend on cloud infrastructure and an active internet connection.

Its capabilities enable it to plan and complete multi-step tasks, call software tools, work with text alongside images, write and debug code, and recover when an action fails. Muse Glimmer also marks a milestone as the first model in the Muse family released with open weights under an Apache 2.0 licence, allowing developers to download and customise it for their own applications.

Making a 30B model fit on local hardware

Running a model of this size locally requires substantial memory, with Meta estimating that Muse Glimmer needs more than 55 GB at full precision. This requirement far exceeds the memory available on consumer graphics cards. To solve this, the company compressed the model weights to approximately 4-bit precision, bringing the language model itself below 20 GB. This reduction leaves necessary memory for response generation data, the image-processing system, and other core components.

Meta designed the complete setup to operate within a 24 GB or 32 GB memory envelope, making local deployment feasible on select high-end personal computers. However, these requirements remain significantly higher than those of standard laptops or desktops. The company tested a compressed version of roughly 17 GB on MacBook M4 Max and M5 Max systems, as well as an NVIDIA RTX 5090.

While compression can sometimes impair accuracy, internal evaluations revealed minimal or no degradation across tested agent tasks. Meta notes that these conclusions rely on its own testing.

Designed to complete tasks across several steps

Meta trained Muse Glimmer specifically for agent-based applications where an AI system must perform a sequence of actions rather than answer a single prompt. For example, a local assistant managing schedules, drafting messages, or organising files must interpret instructions, access various tools, retain context across long tasks, and adjust when an action fails. Muse Glimmer targets these exact workflows, with Meta highlighting end-to-end task completion, tool usage, multi-step reasoning, and failure recovery as core strengths.

The model also uses a dedicated system to process images alongside text, enabling applications to handle screenshots, charts, and documents across more than 100 languages. Benchmarks across agent, coding, reasoning, and multimodal tests, including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, show that it performs strongly against models of similar scale.

Faster generation and open deployment options

To speed up generation, Meta included a smaller companion model based on DFlash that uses speculative decoding. This technique allows the smaller model to propose groups of tokens for Muse Glimmer to check in parallel, rather than generating every token individually. The process reduces latency during extended reasoning sequences and multi-step tasks, preventing repetitive pauses that could make local assistants cumbersome.

Developers can download Muse Glimmer weights via Hugging Face, where Meta provides full-precision and compressed versions alongside documentation and support for custom agent configurations. Integrations for llama.cpp, MLX, and ExecuTorch are also in development. Additionally, the model works with platforms such as Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter.

These options give developers the choice to run the model directly on local hardware or deploy it through existing AI infrastructure. For local deployment, however, the 24 GB to 32 GB memory target remains a key constraint for consumer hardware.