Earnings
Home›Earnings›Analyst Ratings›Meta releases Muse Glimmer, aiming to run AI offline
Meta releases Muse Glimmer, aiming to run AI offline
The open-weight model is built with 30 billion parameters and can fit in under 20 GB of memory using 4-bit quantization, according to the release details.
Meta Platforms released Muse Glimmer, an open-weight AI agentic model designed to run entirely offline on standard consumer hardware, with the company positioning the move as a shift in how the AI stack is delivered.
The model uses 30 billion parameters and is optimized through 4-bit quantization to shrink its memory footprint to under 20 GB, the company’s technical specifications indicate. Muse Glimmer also includes a speculative decoding drafter to accelerate text generation.
In its release framing, Meta said the approach is meant to democratize access by making a high-capability foundation model available under an Apache 2.0 license, rather than limiting deployment to proprietary, closed ecosystems.
Meta’s broader strategy described in the release ties into using developer uptake to stress-test the model and surface optimizations that can later be integrated into its advertising infrastructure, the outlet said.