Meta releases Muse Glimmer, an open-weight AI model that runs locally on laptops

Meta has released Muse Glimmer, a 30-billion-parameter artificial intelligence model designed to run locally on consumer computers.

The company released the model’s weights under the Apache 2.0 licence, allowing developers to download, modify and use it. It is the first model from Meta Superintelligence Labs to be released with open weights.

How Meta’s Muse Glimmer works

Muse Glimmer is designed for AI agents that can handle multi-step tasks without requiring users to guide them through every stage. It can plan tasks, use tools, check results and recover when a step fails.

Meta said the model can run on devices with 24GB or 32GB of memory, including Macs with M4 or M5 chips, as well as computers with a single consumer graphics processing unit (GPU), such as the NVIDIA RTX 5090 or AMD Radeon AI PRO. It does not require an internet connection to function.

The model can also process text and images and has been trained on data spanning more than 100 languages.

Read also: Meta Unveils Enhanced Safety Tools for Nigerian Teens and Parents at the Youth Summit in Abuja

Meta said it used a distillation process based on its larger Muse Spark model to train Glimmer, alongside additional training focused on reasoning, coding and agentic tasks.

The company has released full-precision and 4-bit versions of the model, as well as inference code and a DFlash speculative decoding system designed to speed up text generation. Meta said the DFlash system increased Glimmer’s generation speed by 3.1 times on an NVIDIA RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max, compared with standard generation.

Muse Glimmer is available through Hugging Face, while support for tools including Ollama, LM Studio, llama.cpp and MLX is being rolled out. Meta said it is also working with partners including AMD, Arm, Dell, Intel and NVIDIA to optimise performance across devices.

Meta said developers can use the model for local coding, tool calling, file management and other agentic applications. Meta is also expected to release the weights for Muse Spark 1.2, one of its leading foundation models, in the coming weeks.