Meta Unveils Muse Spark Open: 816B Parameter Local Mac AI Model

Overview

The artificial intelligence landscape is shifting toward extreme scale coupled with local hardware optimization. Meta's latest release, Muse Spark Open, shatters traditional boundaries by introducing an open-weight foundation model boasting 816 billion parameters and an expansive 1-million token context window. Engineered specifically to tackle advanced coding tasks and autonomous agentic workflows, this model introduces a novel paradigm: bringing massive-scale intelligence directly to consumer-grade or prosumer desktop workstations like Apple's Mac ecosystem.

Architectural Scale and Multimodal Foundations

At its core, Muse Spark Open is natively multimodal, designed to process complex inputs across multiple data modalities simultaneously. The inclusion of a 1-million token context window allows developers to feed entire codebases, comprehensive documentation, or hours of video data into the system without hitting traditional memory bottlenecks.

Evaluation benchmarks published alongside the launch position Muse Spark Open as a leading contender among domestic open-weight systems. Its sheer capacity—driven by 816 billion parameters—demands immense computational throughput, yet its structural design allows it to interface seamlessly with specialized desktop environments.

Local Execution on Apple Hardware

One of the most striking aspects of the release is its capability to run locally on Apple's desktop hardware. While cloud deployment remains a primary option via various API platforms, local execution provides unprecedented data privacy, reduced latency, and freedom from third-party server dependencies.

However, running an 816-billion parameter model locally comes with rigorous hardware prerequisites. Operating the system smoothly requires substantial memory infrastructure—specifically, at least 1 terabyte of unified RAM. Meta notes that the model achieves peak performance when deployed on a Mac Studio equipped with the M5 Ultra chip, highlighting the rapid convergence of high-end silicon and massive large language model architectures.

Agentic Workflows and Desktop Automation

Beyond raw inference, Muse Spark Open was built to act. Meta integrated the model with Muse Code, its proprietary AI coding terminal, deploying it across complex agentic workflows. During stress tests, the system successfully managed end-to-end software engineering tasks for a Mastodon client, executing bug fixes, developing new features, and overseeing the full build process.

Furthermore, testing extended into creative software engineering. When paired with DaVinci Resolve, the professional video editing suite, the model demonstrated the ability to autonomously review, analyze, and crop video footage. This showcases a tangible leap in how large foundation models can control local desktop applications to perform multi-step labor.

The Strategic Value of Open-Weight Systems

Releasing such a powerful system under an open-weight license underscores Meta's ongoing commitment to democratizing AI research. Unlike closed, proprietary APIs, open-weight models grant enterprise teams and independent researchers the flexibility to audit, fine-tune, and deploy models securely within custom infrastructures. This transparency not only fosters rapid innovation but also drives cost efficiencies for organizations looking to scale agentic automation without recurring cloud compute penalties.