Meta Platforms CEO Mark Zuckerberg has advocated for reduced U.S. barriers on open-source AI to enhance competition with Chinese counterparts. The tech company introduced a smaller, more customizable model named Muse Glimmer, geared towards running tasks on standard devices like Mac or PC with a single graphics card. Zuckerberg mentioned plans for even larger models in the pipeline, emphasizing the importance of democratizing AI access.
Embracing open-weight AI has gained momentum as companies seek to cut costs and address security concerns following incidents involving models from Anthropic, OpenAI, and Meta. Open-weight models differ from closed models as they allow users to customize their core components, known as weights, for specific needs, offering a cost-effective and flexible alternative.
Recent developments in Chinese open-weight models, such as Kimi K3, have underscored the advantages of open-source AI. Meta, along with prominent tech firms like Nvidia and Microsoft, has expressed support for open-weight AI. Chinese startups have been leading the charge in open-weight models, with offerings like Kimi K3 and Qwen3.8-Max rivaling top U.S. systems, which are predominantly closed weight.
Meta’s shares saw a positive response in premarket trading following the announcement. The company is set to release Muse Spark 1.2, its advanced model, as part of efforts to regain ground in the AI competition. Zuckerberg highlighted the importance of community support amid concerns over data-center expansion, acknowledging the challenges faced by U.S. firms in infrastructure development compared to other countries.
To maintain competitiveness in open-weight models, Zuckerberg stressed the need for policy reforms on data use and distillation methods. Meta intends to establish a governance structure granting independent directors authority over model release safety criteria. The U.S. administration’s decision not to subject open-weight AI models to voluntary safety tests was also noted in the context of the evolving AI landscape.