Meta is shipping model weights again after a year-long drought. Muse Glimmer, a 30-billion-parameter model, dropped on Hugging Face under Apache 2.0, purpose-built to run AI agents locally on a single consumer GPU. It’s the first release from the reorganized Meta Superintelligence Labs, and it marks a direct return to the open-model arena after the Llama 4 debacle.
The context here is a rebuild. Llama 4 missed hard—benchmarks were massaged, the largest variant never materialized, and chief scientist Yann LeCun exited. In response, Zuckerberg poured billions into data provider Scale AI, poached top researchers, and restructured the unit multiple times. Glimmer is the first tangible output from that reset, and a more powerful open-weight model, Muse Spark 1.2, is reportedly weeks away.
On benchmarks, Glimmer holds its own against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, particularly on agentic tasks like tool use and web search. Qwen remains stronger on desktop control and terminal work. As always with vendor-run tests, the numbers need a squint: Meta gathered most comparison data itself and admits its setup isn’t optimized for rival models. Still, the practical achievement is real. Through 4-bit quantization, the model squeezes under 20GB—small enough to run entirely on a MacBook or current consumer GPU, with a helper model accelerating text output by up to 3.1x. The pitch is an agent that handles your calendar, files, and messages without shipping data to the cloud.
What makes this launch more than a product update is the essay Zuckerberg published alongside it. He argues superintelligence shouldn’t concentrate in a few labs, defends distilling other companies’ models as a principle worth protecting, and frames open American models as a strategic counter to Chinese labs. That lands squarely in an ongoing fight with OpenAI and Anthropic, whose CEOs have called distillation theft and pushed for tighter export controls. Zuckerberg’s counter is practical as well as philosophical: Meta trails on frontier capability, but open distribution is where it can plausibly lead. The irony? Just weeks ago, Meta internally restricted its own engineers from using Claude Code and Codex, worried the outputs would taint its training data. That tension isn’t going away.
