Meta released its new AI model Muse Glimmer on August 10, 2026, marking a return to an open-weights release strategy. Developed by Meta's Superintelligence Labs, the dense 30-billion-parameter model is available under the permissive Apache 2.0 license. This launch represents the company's first open-weights release since Llama 4.
Muse Glimmer was designed specifically to run local agent workflows that operate continuously in the background. When 4-bit quantized, the model requires less than 20 gigabytes of VRAM. This allows the system to run on a single consumer graphics card like the Nvidia RTX 5090 or on Mac unified memory. In benchmark tests, the software shows particular strengths in autonomous failure recovery and executing long tool calls.
The release follows a series of product announcements by the technology company. On August 5, 2026, five days prior to Muse Glimmer, Meta presented its proprietary frontier model Muse Spark 1.2 alongside the terminal coding tool Muse Code. Chief executive Mark Zuckerberg simultaneously announced that the weights for Muse Spark 1.2 will also be made open source in the near future. This promise reinforces the company's strategic focus on accessible model weights.
Alongside the model release, Mark Zuckerberg published a 14-page manifesto on Personal Superintelligence on August 10, 2026. In the 6,500-word open letter, the founder called on lawmakers in Washington to provide regulatory relief for training open-source AI within the United States. He also announced the Future Is For Everyone Fund to support local communities near Meta data centers. Additionally, Meta is establishing an independent safety board to review criteria for future model releases.
The availability of compact open-weights models is also driving the ecosystem for local autonomous assistants. Open-source frameworks like OpenClaw leverage models similar to Muse Glimmer to act as continuous assistants running on personal computers or local servers. They use popular chat platforms such as WhatsApp, Telegram, or Discord as user interfaces while executing background processes through protocols like the Model Context Protocol. This development is supported by novel database compression methods designed to store conversation histories efficiently.