On Monday, Meta Superintelligence Labs (MSL) introduced Muse Glimmer, a model with 30 billion parameters and freely downloadable weights on Hugging Face under the Apache 2.0 license. In a position piece, CEO Mark Zuckerberg argued that superintelligence should be widely distributed, claiming safety emerges from a balance of power and that the most dangerous scenario is leading labs holding back capable models. He criticized rivals for framing alignment as enforcing centralized values, while Meta aligns agents with individual user goals. This argument contrasts with Meta’s own Muse Spark, which shipped without open weights in April 2026 as a hosted product. In July, Muse Spark 1.1 launched with the first paid Meta Model API, starting at 1.25 US dollars per million input tokens and 4.25 dollars per million output tokens. Version 1.2 followed in early August with Muse Code, also without downloadable weights. This came after Llama 4 was seen as a disappointment, and chief scientist Yann LeCun left Meta in November 2025. Zuckerberg announced Meta will “soon” resume releasing open source models, with Muse Glimmer as first evidence. Zuckerberg and MSL chief Alexandr Wang signaled open weights for a version of Muse Spark 1.2, with no timeline beyond “in the coming weeks.” Muse Glimmer, not a frontier model, was trained from Muse Spark using logit distillation and post-trained on agentic tasks, processing images via