Meta Releases Open-Weight Muse Spark
The Facts
- Meta released an open-weight version of its Muse Spark AI system on Thursday, letting developers download and adapt a system built for coding and AI agents. The company describes it as "the most capable U.S. open-weight model to date."
- Muse Spark Open carries 816 billion total parameters with 42 billion active at a time and a one-million-token context window. The natively multimodal system accepts text, image, video and audio inputs and returns text, according to its system card.
- The system is optimized to run on two Apple M5 Ultra Mac Studios using the MLX framework. Meta's model card states that other setups need a GPU with at least 700GB of VRAM for the NVFP4 version and 2TB for the BF16 version.
Sources Split
LEFT
CENTER
RIGHT
+2
+5
ESTAB
CRIT
CENTER
PRO
ESTAB
The Spin
Pro-industry narrative
Open-weight leadership belongs back in American hands, with this release proving Meta's AI comeback is real. Letting anyone run a frontier-class agentic system on a desktop matters more than any locked-down API for developers. Freedom to download, tune and deploy advanced AI beats renting intelligence from a black box every time.
Industry-critical narrative
Open-weight bragging rights mean little when the license limits commercial use and independent testing puts the system behind cheaper rivals on agentic coding and real-world tasks. Paying premium prices for middling results is a tough sell when rival open systems deliver more intelligence per dollar. Genuine openness means shared data, methods and accountable governance, not a download link.
