Meta returns to open models with Zuckerberg's plan to out-copy China and sell compute by auction

Why it matters

Meta has re-entered the open-weight model arena with Muse Glimmer, its first openly released model since Llama 4, offered under an Apache 2.0 license on Hugging Face. The compact agent model, distilled from the larger Muse Spark, is designed to run continuously on consumer hardware, keeping personal data on the device rather than in the cloud. The release follows a rocky period: Llama 4 drew criticism over benchmark claims, Meta reorganized its AI efforts into Meta Superintelligence Labs, invested heavily in Scale AI and recruited top researchers, and lost chief scientist Yann LeCun. Mark Zuckerberg paired the release with an essay defending open distribution of powerful AI and explicitly defending distillation, the practice of training on other labs' model outputs, which OpenAI and Anthropic have condemned. Meta compares Glimmer favorably to Google's Gemma4-31B and Alibaba's Qwen3.6-27B on agent tasks, though it concedes ground on desktop and terminal work. The move also highlights investor unease over Meta's massive spending without matching revenue.