Meta's New Open AI Model Arrives with a Pitch for Learning from Competitor Data and a $600 Billion Question

Meta has released Muse Glimmer, a 30-billion-parameter model with weights open under Apache-2.0 on Hugging Face, aimed at AI agents running locally on a Mac or PC with a single consumer GPU. It is Meta's first open model since Llama 4 in spring 2025, and the company plans to release an open-weight version of its strongest model, Muse Spark 1.2, in the coming weeks. During the hiatus, Meta reorganized its AI division as Meta Superintelligence Labs after Llama 4 disappointed due to inflated benchmarks, hired top researchers, invested in Scale AI, and lost chief scientist Yann LeCun. Glimmer is competitive with Google's Gemma4-31B and Alibaba's Qwen3.6-27B, leading in most benchmarks especially agentic tasks, though Qwen is better at desktop control and terminal tasks. The model is compressed to about 4 bits to fit under 20 GB for consumer GPUs and MacBooks, with a small auxiliary model speeding up text output by up to 3.1 times. Zuckerberg's essay 'The Future is for Everyone' argues for broad distribution of superintelligence and defends distillation of competitor models, a practice criticized by OpenAI and Anthropic, who see it as free riding. He calls for fewer restrictions on training data and offers early model access for security checks. Meta's argument is partly business-driven, as it lags behind OpenAI and Anthropic in frontier models. Investors question how Meta will justify $600 billion in spending through 2028, as open models do not generate license revenue; Zuckerberg hints at a dynamic auction mechanism for compute resources, similar to its ad business, but details are sketchy.