Meta releases Muse Glimmer, a 30-billion-parameter open agentic model for local hardware

Meta Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter open agentic model available under the Apache 2.0 license. Designed to operate on consumer hardware, the model allows developers to run autonomous agents locally on Macs and PCs equipped with a single consumer GPU. Weights are currently available on Hugging Face, marking a continuation of Meta’s strategy to distribute fundamental AI research to the broader developer community.

The release addresses the infrastructure constraints that typically limit local AI deployment. While a 30-billion-parameter model at full precision would require more than 55GB of memory, Meta has applied quantisation techniques to compress the weights to approximately 4-bit precision. This reduces the memory footprint to under 20GB, leaving sufficient headroom for working memory and perception encoders within the 24GB or 32GB memory envelopes common in consumer devices.

To ensure practical performance, Muse Glimmer incorporates a DFlash-based drafter that accelerates token generation. This speculative decoding mechanism allows the model to propose entire blocks of tokens simultaneously, which the main model then verifies in parallel. This approach significantly reduces latency compared to standard token-by-token generation, enabling fluid conversation and real-time agent interaction without sacrificing output quality.

The model is optimised for specific agentic workflows, including local coding, function calling, and autonomous agent behaviour. Meta evaluated Muse Glimmer against competitors such as Gemma4-31B and Qwen3.6-27B, noting strong performance within its size category. The architecture supports long-horizon execution, precise tool calling, and long-context memory, which are critical for agents that manage personal context and execute multi-step tasks.

Integration with existing developer tools is a key component of the release. Optimised support for llama.cpp, MLX, and ExecuTorch is scheduled to arrive in the coming days. Meta is also collaborating with industry partners including AMD, Arm, Dell, Intel, and NVIDIA to optimise performance across various devices. Deployment options are further expanded through partnerships with Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter.

Developers can customise the model for specific use cases using PyTorch’s TorchTitan training feature. Meta has also released documentation to assist with setting up custom scaffolds and building agents responsibly. The release underscores the growing viability of smaller, effectively trained models that can approach frontier-level performance on targeted tasks while operating entirely offline.