Meta launches Muse Glimmer, a practical on-device AI model for local workflows

Meta has introduced Muse Glimmer, a 30-billion-parameter model designed for on-device AI tasks, enabling capabilities such as tool use and multi-step reasoning on consumer hardware without relying on cloud-based infrastructure.

Meta has unveiled Muse Glimmer, a 30-billion-parameter model from Meta Superintelligence Labs that is intended for always-on local agent workflows rather than cloud-dependent use. According to Meta, the model is designed to run on a Mac or PC with a single consumer GPU, making it a more practical option for on-device AI tasks such as tool use, coding assistance and function calling.

The company said the model is built for agentic task completion, multi-step reasoning, multimodal input and reasoning, error diagnosis and retry handling. Meta also said Muse Glimmer can operate with or without an internet connection, a notable point as many AI applications still rely heavily on remote servers and network access.

Memory requirements shape the model’s positioning. At full precision, Meta said Muse Glimmer would need more than 55GB of memory, but 4-bit quantisation reduces the footprint to under 20GB, which the company says makes it suitable for 24GB or 32GB systems. Meta also claimed comparative benchmark gains against Google’s Gemma 4-31B and Alibaba’s Qwen3.6-27B, saying Muse Glimmer led in 12 tests, although those results come from Meta’s own evaluation.

Meta said the weights are available under the Apache 2.0 licence and can be downloaded from Hugging Face. The company said support will arrive soon through tools including Ollama, LM Studio and Unsloth, while optimised integrations for llama.cpp, MLX and ExecuTorch are also due. Meta said it is working with AMD, Arm, Dell, Intel and Nvidia to improve performance across devices, and added that the model was trained on data spanning more than 100 languages.