Meta releases Muse Glimmer for local AI agents

Meta has released Muse Glimmer, a 30-billion-parameter open agentic model designed to run locally on a Mac or PC with a single consumer GPU. Meta Superintelligence Labs published the announcement on August 10. The model weights are available on Hugging Face under the permissive Apache 2.0 license, making this a concrete release for developers rather than a product preview. Meta positions it for local agents, function calling, coding and LLM-as-a-judge evaluation.

The release addresses a practical limit in today’s AI market. Most capable agents still depend on cloud servers, network access and a provider’s pricing and availability. Meta says Muse Glimmer can work on a local device, including when there is no internet connection. That could matter for privacy-sensitive workflows, offline use and teams that want more control over their infrastructure.

At 30 billion parameters, the model would normally need more than 55 GB of memory at full precision. Meta says it compresses the weights to approximately four-bit precision, bringing the language model below 20 GB. That leaves room for the model’s working memory, its image-perception encoder and a lightweight DFlash drafter inside a 24 GB or 32 GB hardware envelope. The drafter proposes blocks of tokens that the main model verifies in parallel, a technique intended to make longer agent loops feel faster.

Muse Glimmer is built for more than chat. Meta describes long-horizon task completion, precise tool calls, multi-step reasoning, failure recovery and text-and-image input. The model was trained through distillation from the larger Muse Spark, followed by longer-context agent data, supervised tuning and reinforcement learning. Developer documentation is available now; integrations with tools such as llama.cpp, MLX and ExecuTorch are scheduled to follow, while partners including Ollama, LM Studio and vLLM are preparing support.

The announcement is important for AI makers because capability is moving closer to the user’s hardware. A local agent can reduce recurring inference costs and keep sensitive context away from a third-party server, but it also shifts responsibility for updates, access controls, model safety and performance to the operator. Meta’s benchmark comparisons are its own evaluations, so independent testing still matters. For businesses, Muse Glimmer is a signal to compare cloud and local deployment on total cost, latency, privacy and reliability. For users, the real test is whether a model that fits on a personal machine can remain useful and predictable across everyday work.