Overview
- Meta publicly published Muse Glimmer on Monday as a 30‑billion‑parameter model with weights under the permissive Apache 2.0 license.
- The company says Muse Glimmer uses 4‑bit quantization so it can run entirely on a single 24GB or 32GB GPU, enabling always‑on local agent workflows.
- Meta reported tests on MacBook M4 Max, M5 Max and an RTX 5090 that showed smooth conversation and real‑time agent interaction with computation performed locally.
- CEO Mark Zuckerberg urged U.S. policymakers to ease rules on model training, data access, model use to help American labs compete and announced a $1 billion fund for community adoption.
- Open weights and single‑GPU execution lower costs and broaden access but also raise familiar safety, export and data provenance concerns as Chinese open models intensify competition.