Meta's lightweight Muse Glimmer AI now runs on a single computer

Meta has introduced a new, streamlined AI model called Muse Glimmer that is light enough to operate on a single computer, according to an announcement from the company. Built on Meta's closed Spark 1.2 model, Muse Glimmer requires just one GPU for tasks like scheduling and file management. Meta said it designed the model to balance performance with the memory and processing limits of local hardware.

The model's weights - the settings AI uses to generate responses - are now freely available on Hugging Face, along with developer documentation. Users can download and run the model on their own PCs. Meta also noted that optimized support will soon appear on llama.cpp and other platforms, allowing users to move from download to a working agent in minutes.

Despite its smaller size, Meta says Muse Glimmer performs well on benchmarks like DeepSearch QA, MCP-Atlas, and SWE-Bench, which tests coding and debugging. The model supports reliable tool use, multi-step reasoning, recovery from errors, multimodal input, and works with orchestrators like OpenClaw. It was trained on data from over 100 languages.

The release echoes efforts from Chinese competitors like DeepSeek, which also offer locally runnable models with open licensing. Meta CEO Mark Zuckerberg wrote in an accompanying essay that rather than concentrating superintelligence, the goal should be to distribute it widely.

Meta's Muse Spark AI has been viewed as less powerful than rivals from OpenAI and Anthropic. This new model appears to push Meta toward an open-source approach, appealing to users who prefer running AI on their own machines. The topic is timely, as NVIDIA recently launched the Open Secure AI Alliance to strengthen cyber defense after a rogue, closed OpenAI model attacked Hugging Face.

Source: Engadget