Reinforcing his position in the open-versus-closed AI debate, Meta CEO Mark Zuckerberg on Monday warned against the extreme concentration of AI and called for a rethink of US policies around training data and AI distillation, as the company released the weights of a new open-weight model, Muse Glimmer.
In a 14-page essay titled ‘The Future is for Everyone’, Zuckerberg said that restricting the development and use of open AI models would hand an advantage to Chinese developers, which have emerged as strong contenders in the open-weight AI race. “The notion AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he said, adding that AI should be distributed widely rather than controlled by a small group of companies or experts.
Agentic Workflows
Meta’s latest model, Muse Glimmer, is a 30-billion-parameter model designed to perform agentic tasks and run locally on a laptop or a single consumer GPU. Unlike conventional AI models that largely respond to prompts, Glimmer is designed to plan a task, use developer-approved tools, check the results and recover when a step fails. It can operate within a 24-32 GB memory envelope, with Meta targeting systems including Mac computers powered by M4 or M5 chips and Nvidia’s RTX 5090.
Defence of AI Distillation
The release is also significant because Glimmer was distilled from Meta’s larger Muse Spark 1.2 model. Distillation, which involves using the outputs of a more powerful model to train a smaller one, has become increasingly contentious as Chinese developers have released open-weight models that have narrowed the performance gap with leading US systems. Zuckerberg, however, explicitly defended the technique, stating that restrictions on it could undermine the US position in open AI.
“For the US to lead in open source, we will need to rethink our policies in several areas, including distillation and data use in training,” Zuckerberg said. “Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe.”
He also called for fewer restrictions on training data, saying American AI companies face additional constraints compared with foreign developers.
The intervention comes as the competitive landscape in AI is becoming increasingly divided between open-weight and closed models. While OpenAI, Anthropic and Google have largely retained control over the weights of their leading models, Chinese companies including Moonshot, Alibaba and DeepSeek have pushed open-weight systems that can be downloaded and customised by developers. Meta itself has a long history of releasing open models through its Llama series, although it had moved towards a more closed approach with its latest generation of models.
Meta is now preparing to return to the open-weight strategy at the higher end of its model portfolio. The company said it would release the weights of Muse Spark 1.2, its most advanced model, in the coming weeks. The model was developed by Meta’s superintelligence team, which the company formed last year as it sought to close the gap with frontier AI labs.
The move also comes amid growing concern among businesses over the cost of running increasingly large AI models. Open-weight models can allow companies to run and customise models themselves, reducing dependence on expensive cloud-based inference and giving them greater control over data. Glimmer takes that proposition further by enabling a relatively large model to run directly on consumer hardware.