Key facts
- Meta released Muse Glimmer, a 30 billion parameter open-weight AI model.
- The company plans to release weights for its more powerful model, Muse Spark 1.2, soon.
- CEO Mark Zuckerberg published a 6,000-word essay advocating for decentralized AI and open-weight models.
- Zuckerberg argued against the concentration of power in proprietary AI systems due to existential threat concerns.
- Muse Glimmer is designed to run on local machines, offering an alternative to cloud-based AI services.
Meta has announced its intention to focus on open-weight large language models, releasing its new model, Muse Glimmer, and promising to open the weights for its more powerful Muse Spark 1.2 model in the coming weeks. Alongside these announcements, CEO Mark Zuckerberg published a lengthy essay detailing the company’s philosophy on AI systems and governance, aiming to differentiate Meta from competitors like OpenAI and Anthropic, which develop proprietary models.
Muse Glimmer is a 30 billion parameter model with a 128,000-token context window, designed to run on users’ local machines. This approach contrasts with cloud-based services and APIs. The model's weights are open source under the Apache 2.0 license. Muse Spark, launched earlier this year, was initially a closed, proprietary model, but Meta introduced a paid service with Muse Spark 1.1 and later Muse Spark 1.2 on August 5, accompanied by Muse Code, a terminal coding agent.
Zuckerberg's essay argues for decentralized AI systems and directly addresses concerns about existential threats and misalignment often cited to justify tightly controlled, proprietary AI models. He contends that a singular superintelligence would inevitably prioritize some values over others, making it incapable of being benevolent to everyone. Instead, Meta advocates for models personalized to individual or group needs and values, suggesting that decentralization will enhance safety by distributing the benefits of advanced AI equally.
Meta has been perceived as trailing other major tech companies in foundation model development, with less adoption compared to OpenAI and Anthropic, which have successfully targeted enterprise customers. The company's AI division underwent a significant overhaul last year. The recent advancements in Chinese models like Alibaba's Qwen3.8-Max and Moonshot's Kimi K3, which rival frontier models, have intensified the debate around open-weight models and distillation. Meta appears to be positioning itself as a US-based alternative to these Chinese offerings, emphasizing openness, customizability, and affordability, with a focus on personal use rather than large-scale enterprise deployments.