Meta is making another major push into open AI, this time focusing on powerful AI agents that can operate directly on users’ computers. The company has released the weights of Muse Glimmer, a 30-billion-parameter AI model designed to run locally on consumer hardware. The move could allow developers to create AI agents without depending entirely on cloud-based servers.
Meta CEO Mark Zuckerberg highlighted the release in a post on X, describing Meta as a “strong supporter of open source” and saying he was proud of the launch. He also confirmed that Meta intends to release the weights of Muse Spark 1.2, its newest foundation model, in the near future.
A 30B MODEL DESIGNED TO RUN LOCALLY
One of Muse Glimmer’s biggest features is its 30-billion-parameter size and its focus on local AI agents. Put simply, developers can potentially run the model on compatible Macs and PCs instead of sending every request to a remote data centre.
Meta says Muse Glimmer can operate on a system equipped with a single consumer GPU. The company has also compressed the model to around 4-bit precision, reducing its size to less than 20GB. That makes it small enough to fit within the memory capacity of some high-end consumer devices.
In effect, the model brings AI processing closer to the user rather than requiring constant cloud connectivity.
Local AI models could be particularly useful for applications where privacy, offline functionality and control over data are important. Developers could also create assistants capable of interacting with files, coding environments and other applications while keeping much of the processing on the user’s device.
MORE THAN JUST A CHATBOT
Muse Glimmer is designed to go beyond simply answering questions. Meta describes it as an agentic model, meaning it is built to handle tasks involving multiple steps.
The model can reportedly use tools, maintain plans, work with images, generate and debug code, and recover from errors during a task.
This could enable developers to create AI agents that actually perform tasks rather than merely explaining how users should complete them.
For example, instead of asking an AI how to organise a collection of files, a developer could build an agent capable of examining the files, accessing the necessary tools and carrying out the organisation itself.
META DOUBLES DOWN ON OPEN AI
The release also represents another important step in Meta’s broader artificial intelligence strategy.
The company has built much of its AI approach around open-weight models, particularly its Llama family. However, Meta had previously delayed the full release of its most advanced Muse Spark model amid internal discussions surrounding safety.
With Muse Glimmer now being opened up and Muse Spark 1.2 also set for release, Meta appears to be once again embracing the idea that developers should have direct access to its AI models and be able to build on top of them instead of relying exclusively on closed cloud-based systems.
Zuckerberg reinforced that position in his X post by describing Meta as a “strong supporter of open source” and praising the teams responsible for the models.
The release also aligns with Zuckerberg’s broader views about the future of artificial intelligence. In a separate essay, the Meta CEO argued that superintelligent AI should not be controlled solely by governments or large companies.
That philosophy places Meta’s approach in contrast with the more closed AI strategies followed by companies such as OpenAI, Anthropic and Google.
