Meta releases open-source Muse Glimmer model for local computing

Meta has released Muse Glimmer, a new open-source artificial intelligence model designed to operate on local consumer hardware. The 30-billion-parameter system, which is based on the company’s Spark 1.2 architecture, is optimised to run on a single computer equipped with just one graphics processing unit (GPU). The model weights are available for free on Hugging Face under the Apache 2.0 license, alongside comprehensive developer documentation.

According to Meta, the model was engineered to balance capability against the memory and compute constraints of local devices. It supports agent-oriented tasks including scheduling, file management, code debugging, and multi-step reasoning. The system also features long-horizon execution, precise tool calling, and long-context memory, making it suitable for autonomous agent behaviour on Macs and PCs.

CEO Mark Zuckerberg stated in an accompanying essay that the release reflects a strategic shift to distribute AI capabilities widely rather than centralising them. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," Zuckerberg wrote. The move positions Meta to compete with other providers offering models that can run locally, such as DeepSeek, amid growing interest in secure, on-device computing.

Meta claims the model demonstrates strong performance relative to its size, citing success rates on benchmarks including DeepSearch QA, MCP-Atlas, and SWE-Bench. The system was trained on data from over 100 languages and supports multimodal input. It also offers scaffold compatibility with agent orchestrators such as OpenClaw, allowing for reliable tool use and failure recovery in complex workflows.

The release follows recent developments in the local AI sector, including NVIDIA’s launch of the Open Secure AI Alliance to enhance cyber defence. This initiative comes after a rogue attack on Hugging Face by an unreleased, closed model from OpenAI. Meta’s Superintelligence Labs has previously released models under open licenses, marking a continuation of its strategy to attract users who prefer running models locally rather than relying on centralised cloud services.