Meta released Muse Glimmer, a 30 billion parameter open agentic model built to run locally on phones and laptops, under an Apache 2.0 license, according to a company blog post Monday.
The model handles multi step tasks, tool use and failure recovery, and processes text, images and video across more than 100 languages, Meta said. It is optimized for what the company calls always on local agent workflows: scheduling, drafting messages, organizing files and generating code, either with an internet connection or without one.
Quantized versions of the model fit in under 20 gigabytes of memory, making Muse Glimmer usable on consumer GPUs with 24 or 32 gigabytes of capacity, Meta said. The company reported that quantization and speculative decoding make the model 3.1 times faster on Nvidia's RTX 5090 and 1.8 times faster on Apple's M5 Max MacBooks, and said Muse Glimmer performs strongly against Google's Gemma4 31B and Alibaba's Qwen3.6 27B on agentic and reasoning benchmarks. Weights are on Hugging Face, with integrations planned for llama.cpp, MLX and ExecuTorch.
The release lands alongside a roughly 6,500 word letter Mark Zuckerberg published Monday laying out his vision for what he calls personal superintelligence, TechCrunch reported. Zuckerberg argued that distributing superintelligence widely "has the potential to begin a new era of personal empowerment," describing agents that "will work 24/7 on your behalf" and saying "everyone will have free or affordable access to these tools," according to TechCrunch's account of the letter.
TechCrunch noted a tension in that pitch: Meta draws a line between open models like Glimmer, which anyone can download and modify, and its more capable closed weight Muse Spark model, which stays under company control.
For builders, an Apache 2.0 model that fits on a single consumer GPU and runs offline is a real deployment option for agent workloads that would otherwise need a cloud inference bill or a network connection. Just don't mistake it for Meta's best model. That one isn't part of the open release.