Meta has released Muse Glimmer, a scaled-down "open source" AI model small enough to run on a single GPU, marking a shift towards making its artificial intelligence more accessible for local use. Based on Meta's closed Spark 1.2 model, Muse Glimmer is designed for agent-oriented tasks such as scheduling and file management, and comes as Meta's flagship Spark AI is widely regarded as trailing rivals from OpenAI and Anthropic. The move appears aimed at winning over users who prefer running AI locally rather than relying on cloud-based systems.

Meta is providing the model's weights and developer documentation free via Hugging Face, with optimised integrations planned for llama.cpp and similar platforms to speed up deployment. The company says the model performs well for its size on benchmarks including DeepSearch QA, MCP-Atlas and SWE-Bench, and supports multi-step reasoning, tool use and multimodal input, having been trained on data from more than 100 languages. Meta's approach echoes open-licensing strategies used by Chinese rivals such as DeepSeek, and comes as the wider industry grapples with AI security concerns, including NVIDIA's recent launch of the Open Secure AI Alliance following a rogue attack on Hugging Face by an unreleased OpenAI model.