AI Model Customization: Why it Matters for Policy Makers
BY ASHLEY NAGEL
On August 10, Meta released Muse Glimmer, a 30-billion parameter¹ open-weight AI model designed for agentic (highly autonomous) workloads and running on local hardware. Its smaller size may make it less capable than the largest frontier systems, but also more practical and economical for institutions that want to integrate AI into their operations safely and inexpensively. Meta published the weights under the permissive Apache 2.0 license, which allows anyone to download, modify, and self-host it, and one of the model’s distinguishing characteristics is that it is small enough to run on your personal higher-end Mac or PC. In an accompanying post, CEO Mark Zuckerberg framed the release in terms of individual empowerment, but others see it as a competitive strategy instead. Glimmer may be a classic case of “commoditizing your complement”: a free complement that drives up demand for advertising and devices that Meta primarily profits from, while undercutting competitors whose entire business is charging for model access.
On July 16, Beijing-based Moonshot AI launched Kimi K3, a model with roughly 2.8 trillion parameters, likely the largest “open-weight” model ever released. Independent evaluations indicated that K3 is especially strong at frontend development — building the visual, interactive parts of websites. (Though, it's weaker on cyber. A joint UK-US government evaluation found it performs well below leading models at hacking-related tasks.) K3 delivers highly competitive coding and agentic performance at a fraction of the cost of leading closed models. The release rattled markets, boosted Moonshot's revenue, and reignited discussion in Washington over whether and how the US should respond to capable Chinese models that any American company can freely download and run. Reports suggest the administration is weighing options ranging from advisories to restrictions, even as industry voices point out that a ban on downloadable software would be very difficult to enforce.
Just one day earlier, on July 15, Thinking Machines Lab, the startup founded by former OpenAI Chief Technology Officer Mira Murati, released Inkling, the largest US-built open-weight model publicly available and, by independent rankings, the leading open-weight release from an American lab. Thinking Machines has been candid that other models score higher on raw capability. The company designed Inkling as a strong, well-rounded foundation that organizations can download and adapt with their own data, betting that many institutions will value control and customizability over general purpose capabilities.
Together, the three releases matter for the broader AI ecosystem for three reasons:
The strongest freely downloadable models, a space long dominated by Chinese labs like DeepSeek, now include serious American entrants.
They put competitive pressure on the business model of “closed” providers, since companies can now run near-frontier AI on their own infrastructure instead of paying for access. If a competent model can be downloaded free and run on a laptop, value may shift from the model itself to data and user access surfaces.
They highlight an existing policy tension: openness supports competition, research, and institutional control over data, but published weights cannot be recalled, and any built-in safeguards can be modified by whoever downloads them.
For elected officials, staff, or public servants working within legislative institutions and striving to keep pace with AI’s evolving terminology and classifications, POPVOX Foundation has a new resource: Open, Closed, and Customizable: A Primer on Model Selection. This primer walks through key terms like open-source, open-weight, and closed models, and explains how the main approaches to customization — system prompting, retrieval-augmented generating (RAG), fine-tuning, and training from scratch — differ from one another.
As these projects move forward, institutions will need to decide what kind of openness fits their needs, how much customization and data control each choice permits, and who owns responsibility for safety and accuracy. The policy debate now unfolding within legislatures around the world revolves on questions over whether to restrict foreign open models, how to weigh security risks against the benefits of competition, and who is accountable when downloadable models are misused. Understanding why new models differ and keeping pace with the technical terminology is necessary for anyone shaping AI policy.
¹ A “parameter” is one of the many numerical settings inside a model that gets adjusted during training. Collectively, parameters encode everything the model learned. Often, “parameters” and “weights” are used interchangeably, but strictly speaking, weights are a type of parameter.
