Meta’s Muse Spark Open runs on a pair of Apple M5 Ultra Mac Studios, turning a desktop into a terabyte‑scale AI workstation.
META UNLEASHES MUSE SPARK OPEN: 816B‑PARAMETER MODEL RUNS ON DESKTOP MAC STUDIOS
*Meta AI drops the most capable US open‑weight model yet, promising agentic AI on a personal desktop. The move tests the limits of home‑grown supercomputing and raises safety questions.*
Meta AI has dropped a game‑changing model onto the consumer desktop. Muse Spark Open, a 816 billion‑parameter system, is the first US open‑weight model that can run on two Apple M5 Ultra Mac Studios without cloud assistance. The release marks a decisive step toward personal supercomputing, letting developers push agentic AI from data centers onto a desk. Meta frames the launch as a push for universal access, but the move also forces the industry to confront safety, licensing and performance trade‑offs at a scale previously reserved for hyperscale providers.
Model Specs and Capabilities
Muse Spark Open arrives with 816 billion total parameters, of which 42 billion are active, and a 1 million‑token context window. Meta markets it as a natively multimodal model built for coding and agentic work, excelling at long‑horizon tasks, complex software engineering, and multi‑step reasoning. The model is released as open‑weight, the most capable US open‑weight model to date, and its weights are publicly downloadable. By exposing the full parameter set, Meta invites developers to fine‑tune the model for niche applications while retaining the core agentic abilities that power its benchmark performance.
Desktop Supercomputing with Apple M5 Ultra
The model runs on two Apple M5 Ultra Mac Studios using Meta's MLX framework. Each Mac Studio offers up to 512 GB of unified memory; together they hold the entire model in memory. The M5 Ultra’s 1.2 TB/s memory bandwidth can stream the small slice each token needs, effectively turning the pair into a terabyte‑scale supercomputer on a desk. Meta’s internal tests ran full agentic workflows on‑device, including shipping iOS changes to the open‑source Mastodon client Ice Cubes and driving DaVinci Resolve to review and crop footage frame by frame.
Ecosystem and Accessibility
Muse Spark Open is accessible through API platforms Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Model weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. The model runs on inference engines SGLang, vLLM, TensorRT‑LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA silicon. Developers can further customize the model via PyTorch, using TorchTitan for fine‑tuning. Custom auto‑tunable Helion GPU kernels in SGLang deliver 1.03‑3.95× kernel‑level speedups on NVIDIA and AMD GPUs and a 3% end‑to‑end boost on AMD MI350.
Safety, Licensing and Future Impact
Meta pairs the open‑weight release with post‑training safety measures evaluated against its Superintelligence Scaling Framework. The model met all risk thresholds required for release, with detailed results slated for a forthcoming model card and evaluation report. Weights are released under the Muse Model License, which permits free use for many developers while imposing limited commercial terms. By democratizing a model of this scale, Meta signals a shift toward personal superintelligence, but the safety safeguards and licensing constraints will shape how quickly enterprises adopt the technology.
If the desktop truly becomes a supercomputer, the barrier to building sophisticated AI tools collapses. Developers can now iterate on agentic workflows without waiting for remote inference, accelerating innovation across software engineering, media production and beyond. Yet the open‑weight nature also spreads risk, making Meta's safety framework the de‑facto benchmark for future releases. The next wave of AI will be measured not just in FLOPs, but in how safely it can be run on a kitchen table.
Sources: Hacker News article linking to Meta AI blog, Hugging Face model page, 9to5Mac coverage.