Meta launches a small AI model that can run locally on a single graphics card; Zuckerberg calls on the US government to lower entry barriers for open-source AI.

Meta CEO Mark Zuckerberg announced on Monday that the company will soon resume releasing some open-source AI models and will launch a new lightweight model, Muse Glimmer, while urging the US government to reduce policy restrictions on open-source AI.

The new model, named Muse Glimmer, is a streamlined version of Meta’s Muse Spark 1.2 model, featuring 30 billion parameters and capable of running on a single graphics card. It is primarily designed for agent tasks such as schedule management and file organization. The model weights will be available for download on the Hugging Face platform, and Meta also plans to release the weights for more powerful versions of Muse Spark.

On the governance front, Zuckerberg announced that Meta will establish a new governance structure, empowering the company’s independent board of directors to approve model release safety standards, responding to ongoing concerns about the security risks of open-weight models.

In his statement, Zuckerberg directly pointed to structural barriers at the policy level. He noted that US labs face many additional restrictions regarding training data, giving foreign labs several advantages in the open weights domain. “US policy must reduce this extra resistance.” He also emphasized that restricting access to foreign open-source models is not an effective solution.

Chinese startups currently lead the competitive race in open weights models. Moonshot’s Kimi K3, Alibaba’s Qwen3.8-Max, and DeepSeek’s V4-Flash all match the performance of top US AI lab systems. In contrast, the leading models from OpenAI, Anthropic, and Alphabet’s Google are all closed-source.

According to two sources familiar with the discussions cited by Reuters, the Trump administration earlier this month informed AI developers that it would not conduct voluntary safety testing on open-weight AI models—a move that some in the industry see as an indirect endorsement of the open-source approach.

Compared to closed-source models from “frontier labs” such as OpenAI and Anthropic, open weight models are generally less costly and have core components accessible to the public, making customization easier for users. As companies grow increasingly cautious about ballooning AI expenses, and as recent cybersecurity incidents affect several mainstream AI platforms, this model is gaining more and more market acceptance.

Last month, AI programming collaboration platform Hugging Face revealed that it used a Chinese open weight model to respond after a rogue OpenAI model attacked the platform, citing functional limitations of closed-source models in cybersecurity applications. In his statement, Zuckerberg also explicitly advocated for AI model distillation technology—using powerful AI systems to train smaller models—which is the core technological path of Muse Glimmer.

The release of Muse Glimmer reflects Meta’s ongoing capital logic of heavy investment in AI infrastructure. Meta, along with Amazon, Alphabet, and Microsoft, has pledged a combined investment of over $2 trillion to expand AI capabilities, with related expenses covering hardware procurement and data center construction. Meta is also planning a cloud infrastructure business, aiming to follow Amazon Web Services’ model by selling AI computing power and model access to external customers.

Meanwhile, Meta announced on Monday the establishment of a $1 billion fund to invest in US communities hosting its owned and operated data centers. Previously, hyperscale cloud companies—including Meta and Google—have encountered local pushback over resource use when expanding data centers in various US locations. This fund is viewed as a move to ease relationships with local communities.