Meta's 816B open model runs on two Mac Studios

Muse Spark Open is Meta's frontier open-weight multimodal model, featuring 816B total parameters and a 1M-token context window designed specifically for coding and agentic workflows.

Its headline trick is breaking frontier compute out of server farms. Optimized with Apple's MLX framework, it is the first model of this scale configured to run locally on a pair of M5 Ultra Mac Studios.

Squeezing 816B parameters onto a desk

Under the hood, Muse Spark Open uses a sparse Mixture of Experts (MoE) setup. The network spans 62 layers, each housing 256 experts. For any single token, the router only fires up 8 experts.

This means while you get the massive reasoning capacity of an 816B-parameter network, active compute per token sits at around 42B parameters. It runs at the speed of a mid-sized model while keeping a massive knowledge base intact.

Memory capacity remains the primary challenge. Because all 256 experts must live in memory simultaneously, running unquantized weights requires roughly 1.5 TB of storage. On standard cloud hardware, that normally demands a cluster of four 8-way H200 GPU nodes.

Meta bypassed this cluster requirement by tapping Apple Silicon's unified memory. An M5 Ultra Mac Studio packs up to 512 GB of unified memory. By pooling two machines via MLX, you get a 1 TB memory pool running across 1.2 TB/s of bandwidth, effectively turning two desktop computers into a mini supercomputer. Meta also released an NVFP4 4-bit checkpoint that runs on a single Nvidia B200 system.

Real-world performance and community doubts

Input-wise, the model natively accepts text, images, video, and audio, outputting clean text. Audio is converted at 12.5 tokens per second, while video is handled frame-by-frame through a dedicated perception encoder.

In demos paired with Meta's Muse Code terminal, the model handled complex, multi-step engineering. It successfully debugged and compiled Ice Cubes, an open-source Mastodon client, and controlled DaVinci Resolve end-to-end to isolate footage highlights and crop them vertically.

Independent reviewers, however, have voiced skepticism. Early community tests suggest Meta's self-reported benchmarks were heavily cherry-picked. On independent intelligence indices, the model trailed competing open-weight models like GLM 5.3, Step 5, Kimiko Free, and Mimo 2.6. Third-party hosting prices on API platforms are also running relatively high.

Hardware specs and license friction

Getting Muse Spark Open running is neither simple nor friction-free. Aside from needing two top-tier Mac Studios, gaining access requires clearing several bureaucratic hurdles.

To download the weights on Hugging Face, Meta asks for personal details, including your legal name, birth date, and organization. Its Muse Model License also mandates a prominent "built with Muse Spark open" badge in your product UI, alongside custom enterprise licensing if your revenue tops $20 million within 12 months.

If you have the hardware, weights are live on Hugging Face, AWS SageMaker JumpStart, and Google Cloud Model Garden, with inference support across MLX, vLLM, SGLang, and llama.cpp. Managed API endpoints are also available via Groq, Fireworks AI, and Together AI.

Why it matters

For engineering teams with strict IP privacy requirements and top-shelf Mac hardware, Muse Spark Open is a compelling blueprint for offline coding agents. Ingesting an entire codebase across 1M tokens locally without sending a single packet outside your firewall is a legitimate breakthrough.

For everyone else, we think it makes little sense right now. The hardware barrier is punishing, third-party benchmarks show it trailing lighter open alternatives, and the restrictive UI attribution terms make commercial deployment messy. It stands as an impressive proof of concept for desktop supercomputing, but pragmatic builders have better options.