Scott Bessent endorses Meta’s Muse Glimmer as win for innovation

US Treasury Secretary Scott Bessent has thrown his weight behind Meta’s release of Muse Glimmer, a new open-source AI model, calling it a win for American innovation and a bulwark against foreign intellectual property theft. The endorsement ties together two threads that rarely share the same sentence: open-source software philosophy and national security strategy.

Meta launched Muse Glimmer on August 10, 2026, a 30-billion-parameter dense multimodal model built to run locally on consumer-grade GPUs. It ships under the Apache 2.0 license, meaning anyone can use, modify, and distribute it without paying Meta a dime.

What Muse Glimmer actually does

The model is designed for what the AI community calls “agentic workloads,” tasks that run for extended periods, make decisions, and recover from failures without constant human babysitting.

Muse Glimmer supports a context window of more than 120,000 tokens, giving it the ability to hold large amounts of information in working memory during a single session. That context length puts it in competitive territory with several leading commercial models, but with a critical distinction: it runs on hardware people already own.

Meta engineered the model to be compatible with consumer NVIDIA, AMD, and Apple GPUs. Most frontier AI models require data center infrastructure that costs thousands of dollars per hour to rent. Muse Glimmer is positioned as a distilled version of Meta’s larger, closed Muse Spark 1.2 model, trading some raw capability for the ability to run on a desktop machine.

Bessent’s open-source pitch

Bessent’s endorsement went beyond a polite nod. The Treasury Secretary explicitly connected open-source AI development by American companies to the broader goal of protecting US intellectual property from foreign theft.

The geopolitical backdrop matters here. US-China tensions over AI development and intellectual property have intensified steadily, with export controls on advanced chips and growing scrutiny of Chinese AI labs that have been accused of training on proprietary Western datasets. Bessent’s endorsement positions open-source releases not as charity but as competitive strategy.

Meta has been building this playbook for years through its Llama family of large language models, which established the company as the most prominent corporate backer of open-weight AI. Muse Glimmer extends that commitment into multimodal territory, where models process not just text but images, audio, and potentially other data types.

Why the local-first approach matters

Running AI locally instead of through cloud APIs changes the economics and the power dynamics of the technology. When a model lives on your own hardware, your data never leaves your machine. There is no API bill that scales with usage. There is no vendor who can change terms of service or shut off access.

The hardware requirements are still meaningful. A 30-billion-parameter model needs a GPU with substantial memory, likely 24GB or more of VRAM for comfortable inference. That rules out the cheapest consumer cards but sits well within the range of enthusiast-grade hardware from NVIDIA, AMD, and Apple’s M-series chips.

The Apache 2.0 license is also a strategic choice worth noting. Unlike some of Meta’s previous Llama releases, which carried custom licenses with usage restrictions for companies above certain revenue thresholds, Apache 2.0 is one of the most permissive open-source licenses available. There are essentially no restrictions on commercial use.

Meta’s decision to release a distilled version of its larger proprietary model rather than open-sourcing the full Muse Spark 1.2 follows a pattern familiar in the industry. Companies keep their most capable systems behind closed doors while releasing smaller, more efficient variants that demonstrate their technical leadership and build developer loyalty.

The competitive implications extend beyond Meta. OpenAI has moved toward more closed development. Google has released some open models through its Gemma family but keeps its flagship systems proprietary. Anthropic remains entirely closed-source. Bessent’s public endorsement of the open approach adds a layer of political incentive to what was previously a purely business calculation.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.