Meta、オープンウェイトモデル「Muse Glimmer」を公開し戦略転換

3 Min Read

After more than a year of focusing on closed models, Meta is shifting its strategy back toward open models with the release of Muse Glimmer, as enterprises look at ways to be closer to their data and maintain greater control over infrastructure.

Released on Monday under the Apache 2.0 license by Meta Superintelligence Labs, Muse Glimmer is a 30-billion-parameter model optimized for local, agentic workflows. It can run on a Mac or PC with a single GPU. The model is designed for complex, multi-step agentic workloads and tasks such as coding, web research and debugging.

Meta launched Glimmer a week after it introduced Muse Spark 1.1, a closed model for advanced reasoning and complex agentic tasks. Glimmer indicates that Meta plans to remain competitive in the open model market as interest in China’s price and performance-competitive open models grows.

The release is also a return of sorts to Meta’s open model strategy, which began in 2023 with the release of the Llama foundation model. Since then, however, Meta has deemphasized Llama, despite releasing a fourth generation of the model a year ago.

Related:Microsoft Framework to Cut AI Agent Training Costs

A Hybrid Approach and the Edge

“I don’t think Meta is simply switching back to open models,” said Sid Nag, president and chief research officer at Tekonyx. He said that Meta is moving toward a strategy in which it intends its closed models to be competitive with models from the frontier AI labs. Meanwhile, it is fielding open models to maximize its reach and appeal to enterprises, even those outside its social media platform ecosystem, similar to Google’s strategy, in which its open Gemma models complement its Gemini closed model line.

Moreover, while the timing of Meta’s new models appears related to the recent wave of Chinese open models and the competitive threat they pose to U.S. closed models and also U.S. open models, Glimmer is different from models from China-based AI vendors Moonshot, Alibaba and others.

“This is a very lean, lightweight model that primarily runs on your desktop,” said Arun Chandrasekaran, an analyst at Gartner, referring to Glimmer’s comparatively small size with only 30B parameters. In contrast, some of the popular new Chinese models boast a trillion-plus parameters.

“Meta is going smaller, and Meta is going directly toward the edge of the endpoint devices where the AI is running,” Chandrasekaran continued. “It’s primarily meant to cover more local desktop-bound agentic workflows, rather than like a more centralized cloud-based model.”

He added that Meta paid close attention to how many enterprises are looking to deploy their models, especially with the success of OpenClaw, now under the OpenAI umbrella, and Anthropic Claude Cowork.

Related:Oracle Brings Google Gemini Models to Enterprise Customers

“People want to run these models closest to where the data is, closest to where the workflow is,” Chandrasekaran said. “To that end, Meta wants to be part of that ecosystem, Meta wants to enable that, Meta wants to hopefully empower more workers in the enterprise to run AI more locally.”

Need for Control

Enterprises running models locally also reflect a demand for control, Nag said. He said many enterprises are looking for domain-specific models, which differ from what most frontier model providers can offer, except Anthropic, which has capitalized on Cowork as a way to attract enterprises looking for domain-specific models.

Moreover, open models favor inference economics, Nag said.

“Training a frontier model costs a lot of money, running millions of enterprise inference workloads costs even more over time,” he said. “By releasing weights, Meta lets customers bear much of the infrastructure cost instead of operating every workload themselves.”

However, Meta still has much to do to gain enterprise adoption, Chandrasekaran said.

“It needs to build that platform layer on top of the model because that’s what enterprises want,” he said. “Meta should also provide better capabilities around security and legal indemnification for enterprises.”

Related:Prompt: Enterprise AI Must Prove Its Value Beyond Deployment

About the Author

News Writer, AI Business

Esther Shittu has covered AI technologies and industry trends since 2021. As co-host of the Targeting AI podcast, she talks with experts, thought leaders and practitioners exploring critical AI developments. Before AI Business, she wrote for SearchEnterpriseAI, the New York Daily News, Bklyner and the Brooklyn Daily Eagle. When she's not diving deep into the world of AI, she spends her time on passion projects and raising her three daughters.