Meta is opening the weights of Muse Spark 1.2, its latest AI model. The company is also launching Muse Glimmer, a family of smaller models designed for agentic tasks that can run on a Mac or PC with a single graphics card.
The releases extend Meta’s approach to open-weight AI at a time when OpenAI and Anthropic continue to keep their leading models closed.
CEO Mark Zuckerberg is tying that competition to a policy debate in Washington. He has called for changes to U.S. approaches to training data and distillation, arguing that restrictions on American developers could make it harder for them to compete with foreign alternatives.
Why It Matters: Meta is giving enterprises another U.S.-developed open-weight option as they decide where to run AI workloads and how much of the model stack to manage themselves. Glimmer also expands the range of agentic work that could be handled locally, potentially reducing reliance on cloud inference while raising questions about data access and controls on managed device
- Meta Extends Its Open-Weight Strategy: Meta plans to release the weights for Muse Spark 1.2 as it works to turn its substantial AI spending into more competitive models. The company expects capital spending of up to $145 billion this year, increasing scrutiny of whether that spending can produce competitive AI models.
- More Agentic Work Could Move To Devices: Glimmer could allow more agentic workloads to run locally instead of relying on cloud inference. That could reduce costs and keep certain data on the device. For enterprises, it also means considering what locally running agents can access and how those endpoints are managed.
- Meta Offers An Alternative To Closed Models: Access to the underlying weights gives companies more control over how Muse Spark 1.2 is deployed and operated. That could appeal to enterprises that want a U.S.-developed model they can run in their own environments instead of relying entirely on a provider-operated service.
- Chinese Open-Weight Models Add Competitive Pressure: Alibaba, DeepSeek, and Moonshot have released open-weight models that compete with leading U.S. systems in some areas. Zuckerberg argues that restricting access to foreign open models is not an effective response. His view is that American developers need stronger open models of their own to compete globally.
- Distillation Rules Could Shape Model Development: Zuckerberg is also calling for changes around training data and distillation, where output from a more capable model is used to help train another. The practice has raised intellectual property concerns among some U.S. lawmakers. How those issues are handled could affect companies using model outputs to develop or customize their own AI systems.
