Meta has introduced Muse Glimmer, a 30-billion-parameter open-weight model for local AI agents, while Mark Zuckerberg has outlined plans to resume releasing some open models and explore auction-based access to paid compute.
Meta has released Muse Glimmer, a 30-billion-parameter open-weight model designed for agentic AI workloads, according to a Meta AI Research announcement. The company said the model is intended to support local agents that can write code, call functions and use external tools.
Meta is distributing Muse Glimmer under the Apache 2.0 licence, a permissive software licence that allows broad reuse subject to its terms. The release focuses on systems that can take structured actions across multi-step tasks, rather than solely responding to conversational prompts.
The Decoder described Muse Glimmer as the first open model from Meta’s Superintelligence Labs. It reported that compressed model weights can run in less than 20 GB of memory, potentially allowing the model to operate on consumer hardware. Meta’s own announcement characterizes the release as a model that can run on-device, though practical performance will depend on the compression method and hardware used.
The release coincides with Mark Zuckerberg’s essay, The Future is for Everyone. In the document, Zuckerberg said Meta would resume releasing some open-source AI models and argued that U.S.-developed open models should remain competitive globally.
Zuckerberg’s position is not a commitment to publish every future Meta model. His wording leaves room for the company to offer both open-weight and closed systems. Still, it marks a renewed public emphasis on releasing model weights after a period in which leading AI developers have increasingly kept their most capable systems proprietary.
The Decoder said Zuckerberg’s essay also defended model distillation and called for fewer restrictions on U.S. AI labs. Those arguments contrast with the approach of companies including OpenAI and Anthropic, which have generally retained control of the weights for their most advanced models while emphasizing controlled deployment and safety measures.
Zuckerberg’s essay also proposes a dynamic auction mechanism for paid compute access. Axios reported that the proposal is presented alongside Meta’s plan to release more open models.
The broad idea is that prices for computing capacity would adjust in response to demand, potentially allocating scarce paid AI infrastructure through an auction-style market. Meta’s published material does not specify how such a system would work in practice, which customers could participate, or whether the company intends to launch it commercially.
The Decoder additionally reported, citing The Wall Street Journal, that an open-weight version of Muse Spark 1.2 could follow Muse Glimmer. Meta’s research post, however, specifically confirms the release of Muse Glimmer.
Taken together, the announcements show Meta linking a local-capable open-weight model with a wider argument for decentralizing access to AI models and computing resources. Whether that strategy extends to the company’s most capable future systems remains unclear.
Meta introduces Muse Glimmer Meta has released Muse Glimmer , a 30 billion parameter open weight model designed for agentic AI workloads, according to a Meta AI Research announcement.
The company said the model is intended to support local agents that can write code, call functions and use external tools.
Meta is distributing Muse Glimmer under the Apache 2.0 licence, a permissive software licence that allows broad reuse subject to its terms.
Continue reading
