Meta unveils 'Muse Glimmer' in new push for AI dominance

Meta CEO Mark Zuckerberg has called for relaxing some restrictions on open-weight AI models to help the United States stay ahead of China in artificial intelligence. Meta also released a new open-weight AI model, 'Muse Glimmer'. During the unveiling on Monday, Zuckerberg said the company has plans to launch even larger models in the future.

Although 'Muse Glimmer' is smaller in size compared to the leading models of its competitors, it is capable of performing various tasks automatically using just a single graphics card on a personal computer or Mac. Zuckerberg argued that AI technology should be more widely distributed rather than controlled by just a few companies.

In a 14-page essay, Zuckerberg stated that open-weight AI models are relatively less expensive for businesses and offer the flexibility to customise them according to their needs. China's MoonShot, Alibaba and DeepSeek have already developed such models, competing with the US's top AI systems. In contrast, the main models from OpenAI, Anthropic and Google remain largely closed.

Meta has also announced that it will release the weights for the more advanced 'Muse Spark 1.2' model. The move follows a shift in the company's AI strategy after 'Llama 4' did not receive the expected response. Zuckerberg also announced a $1 billion fund to address local concerns over massive investments in AI infrastructure, aiming to support communities affected by data centre construction.

Zuckerberg claimed that the US is lagging behind China in building AI infrastructure. He said Meta could spend as much as $145 billion on the sector this year. He also called for a review of rules on data usage and existing restrictions on 'distillation' methods to accelerate US companies' progress in open-weight AI models.

Meta also stressed caution on safety. Zuckerberg said that independent directors would be responsible for approving safety standards before any model is released. He also noted that he does not consider restricting access to foreign open-weight AI models to be an effective solution.