Meta is pivoting back toward an open-weight strategy to differentiate itself from proprietary competitors like OpenAI and Anthropic. The company has released Muse Glimmer, a 30-billion parameter model released under the Apache 2.0 license, designed specifically for local inference on consumer GPUs rather than cloud-based APIs. This model features a default context window of 128,000 tokens and was created through distillation from the more powerful, proprietary Muse Spark model. Meta also pledged to release the weights for Muse Spark 1.2 in the coming weeks, following the August 5 release of the model and its companion terminal coding agent, Muse Code. While Muse Code aims to compete on cost against open-weight models from Chinese labs, it has not yet matched the performance of frontier models from Anthropic or OpenAI. This strategic shift follows a period of internal reorganization and a brief departure from open-source principles, including the July release of Muse Spark 1.1 which introduced Meta's first paid service. By releasing Glimmer, Meta seeks to support the movement toward local device inference, reducing industry-wide reliance on massive cloud-hosted models and addressing concerns regarding large-scale model distillation used by competitors.