Meta launches Muse Glimmer to bring powerful AI models to consumer devices

Meta's new Muse Glimmer model focuses on local AI, coding, function calling and AI agents, while giving developers access to its weights under the Apache 2.0 licence.

To run on consumer devices without being dependent on all the data centres in the cloud, Meta has introduced a new open-weight small language model called Muse Glimmer.

Meta unveiled the model on Monday, August 10, as part of its efforts to create an open-sourced AI model that can be used and modified by developers. Muse Glimmer has 30 billion parameters and is optimized for local AI agents, coding, function calling and more.

Muse Glimmer is designed to be run on a Mac or PC using only one consumer GPU, as opposed to many large AI models that depend on data centre infrastructure for powerful processing capabilities. This may facilitate the use of powerful AI models on individual devices, and promote the development of such models by the efforts of individual developers.

Muse Glimmer is licensed under Apache 2.0

Meta has released the weights of Muse Glimmer under the permissive Apache 2.0 licence. The model is available for download by developers from Hugging Face and can be updated or deployed by them to their own applications.

It is created by the AI division of Meta, called Meta Superintelligence Labs (MSL) and led by Alexander Wang. As Meta races against other open-source companies like OpenAI and Anthropic, the company has been putting more resources into open-source AI.

Mark Zuckerberg, the CEO of Facebook, also revealed the new version of Muse Spark 1.2. The weights of the model will be released to the public, he said, prompting developers to be able to download and use those weights.

Zuckerberg also announced the creation of a new family of open models, called Muse Glimmer, designed to be run directly on consumer hardware such as laptops.

Why Meta is focusing on open-weight AI

Meta's move follows a trend of increasing debates on open and closed AI models.

Many organizations like OpenAI and Anthropic have taken the approach of closed models with hidden model weights. In the meantime, some companies, such as DeepSeek, Alibaba and Moonshot, have created open-weight models that can be downloaded and executed by developers.

Zuckerberg thinks that U.S. AI startups should be more aggressive in the open AI space. He has also urged the U.S. to change policies regarding training data and model distillation and distillation.

Zuckerberg's solution to preventing AI power from falling into the hands of only a few companies is to open the models up more widely.

Emphasis on to make the local AI more practical

A key distinction between Muse Glimmer and other bigger AI models is its emphasis on local deployment.

The majority of the most popular AI services are dependent on cloud infrastructure. The requests placed by users are forwarded to the remote servers, where they are processed by extremely strong GPUs and then sent back to the user.

A local run model can help to cut down the need for cloud computing. Depending on the deployment model, it may also be able to provide quicker responses and provide users more control over their data.

Muse Glimmer might thus be beneficial for developers creating AI agents designed to run directly on PCs and other consumer devices.

The 30-billion-parameter model is optimized for coding, function calling and running AI agents. The actual performance will depend on the hardware it is run on and developer optimisation of applications.

Different approach as compared to OpenAI and Anthropic

Meta's approach with open weight functions distinguishes it from other startups like OpenAI and Anthropic.

Closed AI models enable organizations to have more control over their technology, while open-weight models provide developers with even more flexibility in downloading, modifying, and self-hosting the model.

This is especially significant today amid the shift in developers to seek alternatives to cloud-based AI services and businesses to lessen their reliance on expensive AI infrastructure.

The competition has also no longer been restricted to companies in the USA. Chinese AI labs have released various open-weight models, driving a race between the two nations for competitive alternatives. China has seen several open-weight models from the AI labs, pushing the American companies to come up with some alternatives.

Meta's approach is to make their AI models accessible to consumers as well as working on smaller models that can be executed directly on smart devices.

Meta would like some AI abilities to run on the devices people already have, instead of having to process every AI task in a data centre.

This may provide developers some flexibility. For consumers, it may lead to faster, more private AI experiences, without relying on cloud-based services.

The launch also reflects the increasing gap between those companies that develop closed AI models and those that develop open-weight technology. By opening up powerful AI, Meta is clearly hoping to be able to keep up with both the Western AI giants and the fast-growing open-model AI ecosystem that's emerging in China.

Get latest Tech and Auto news from Techlusive on our WhatsApp Channel, Facebook, X (Twitter), Instagram and YouTube.