Meta released Muse Glimmer, a 30-billion-parameter open-weight model designed to run on a single consumer GPU on a Mac or PC, optimized for on-device agentic workflows. The model is released under an Apache 2.0 permissive license via Hugging Face, with integrations for llama.cpp, MLX, and ExecuTorch landing in coming days. Meta also committed to open-sourcing the weights of its flagship Muse Spark 1.2 model.
Muse Glimmer is distilled from Muse Spark using a novel training recipe that transfers agentic reasoning from the larger teacher model while adding quantization and inference optimizations. The model was evaluated against DeepSeek QA, SWE-Bench, and other end-to-end agentic task benchmarks, delivering strong performance in tool use, multi-step reasoning, and failure recovery.
For architects building local-first AI stacks, Muse Glimmer addresses the cost and latency penalties of cloud inference while positioning Meta squarely against Chinese open-weight labs (DeepSeek, Moonshot, Alibaba) dominating the efficient model space. CEO Zuckerberg's accompanying 6,500-word essay directly challenged U.S. policy barriers on distillation and training-data restrictions that he argues handicap American open-source labs versus foreign competitors.