Meta Releases Muse Spark Open: 816B Parameter Model for Macs
Meta partners with Apple to deliver a massive open-weight model capable of running directly on macOS.
Meta has officially launched Muse Spark Open, claiming the title of the most capable open-weight artificial intelligence model available in the United States. In an unprecedented collaboration with Apple, Meta specifically engineered this massive 816-billion-parameter model to execute locally on high-end Apple silicon hardware. This release marks a pivotal moment for developers, privacy-conscious enterprises, and AI researchers who seek frontier-class multimodal performance without relying on closed commercial cloud APIs or surrendering control over proprietary data.
Key Details
Meta’s launch of Muse Spark Open represents a bold leap forward for local foundation model deployment. Featuring 816 billion total parameters and an expansive 1 million token context window, Muse Spark Open brings frontier capabilities directly to desktop workstation environments. Although the model operates natively on macOS, running an open-weight model of this magnitude requires exceptional computational resources and unified memory bandwidth.
To execute Muse Spark Open locally, hardware configurations demand a minimum of 1 terabyte of unified RAM, making top-tier Mac Studio machines equipped with Apple’s M5 Ultra chips the primary hardware target. For software developers and enterprise teams capable of meeting these steep system requirements, the model provides an uncompromised, air-gapped environment for high-throughput coding, video production, and agentic desktop automation.
What This Means
The release of Muse Spark Open signals a fundamental shift in the desktop AI ecosystem. Historically, running frontier-grade AI required routing requests through centralized server farms owned by hyperscale cloud providers. By optimizing an 816B parameter architecture specifically for Apple silicon's unified memory architecture, Meta and Apple are establishing desktop workstations as viable hosts for autonomous agentic workflows.
For organizations handling sensitive intellectual property or operating under strict regulatory compliance frameworks, local execution removes data exfiltration risks inherent in cloud-based APIs. Furthermore, open-weight accessibility guarantees that developers can inspect weight matrices, fine-tune specific behavior, and build custom integration harnesses without fearing sudden API deprecations or restrictive usage policies.
Technical Breakdown
To demonstrate the practical utility of Muse Spark Open across complex workflows, Meta paired the model with Muse Code—its dedicated terminal assistant—and tested it across demanding real-world engineering tasks:
- Massive Context Window: Supports up to 1 million tokens of context, enabling full repository analysis and long-form document synthesis in local memory.
- Native Multimodal Architecture: Processes text, visual assets, and UI states directly without needing separate vision encoders or external API translation layers.
- Autonomous Software Engineering: Successfully executed full build cycles, bug remediation, and feature implementations for complex open-source applications, including a full-featured Mastodon client.
- Creative Software Automation: Demonstrated agentic desktop control within DaVinci Resolve, reviewing raw video footage, identifying key cuts, and executing precise timeline trims.
- Optimized Unified Memory Pipeline: Leverages Apple silicon's unified memory architecture to manage heavy memory bandwidth demands across 816 billion parameters.
Industry Impact
Meta's decision to maintain an open-weight release strategy despite intensifying competition from proprietary labs like OpenAI and Anthropic underscores the growing division in AI distribution. Open-weight models offer unmatched transparency, cost predictability, and customization opportunities for enterprise deployment. By making Muse Spark Open accessible across popular third-party API platforms while supporting local macOS execution, Meta is actively challenging the pricing power of closed API providers.
Moreover, this launch cements Apple's position as a crucial hardware layer for frontier AI execution. While competing hardware vendors push specialized server accelerators, Apple's high-capacity unified memory approach positions its workstation lineup as an essential platform for local, enterprise-grade AI execution.
Looking Ahead
As hardware accelerators advance and quantization techniques improve, the memory footprint required to run 800B+ parameter models locally will continue to contract. The debut of Muse Spark Open on macOS represents early evidence that local desktop environments can support autonomous, frontier-level AI execution. Developers and enterprise engineering teams should expect rapid ecosystem adoption as open-source tools and local serving frameworks optimize for Meta's new flagship architecture.
Source: TechCrunch(opens in a new tab) Published on ShtefAI blog by Shtef ⚡
