Meta's 800B-Parameter Muse Spark Open Runs Locally on Paired Mac Studios
Meta released Muse Spark Open, an 800-billion-parameter open-weight AI model built to run locally on M5 Ultra Mac Studio machines. Engadget
Parameter count is a rough measure of size and capacity. Open-weight means Meta publishes the trained weights, so others can download, run and inspect the model themselves.
Running it locally requires 1TB of memory. The configuration Meta points to joins two 512GB M5 Ultra Mac Studios, and the two machines operate as one system with a terabyte of unified memory, memory shared between processor and graphics.
That memory requirement sets its physical limits. It does not fit on a laptop or a single consumer desktop. It runs as a workstation cluster used as a local inference target, where inference is the step that generates answers from a trained model. Here unified memory does work normally assigned to a rack of accelerators with discrete high-bandwidth memory.
Meta describes Muse Spark Open as the most capable US open-weight model to date. Meta says benchmark results place it ahead of Chinese models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. Agentic tasks test tool use, multi-step planning and tracking state across calls. Visual reasoning tests joint reading of images and text, not just captioning.
The model is available to download from Hugging Face. Teams can pull it directly into existing evaluation and serving pipelines instead of using a hosted API only.
The release continues a line Meta has built through the summer. Muse Glimmer, launched on August 10, 2026, is a 30-billion-parameter open-weight model for always-on local agent workflows and small enough to run on a Mac or PC. Meta Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, arrived with a preview open to developers. Muse Spark was the first text and reasoning AI model from Meta's superintelligence team. Meta has also said it plans to release the weights of Muse Spark 1.2, described as its most advanced model.
The broader context here is the split between Glimmer and Spark Open. One is built for persistent, low-footprint agents on standard client hardware. The other aims for maximum capability in a local package, at the cost of a terabyte of memory and Apple Silicon clustering. For teams building agents, that maps to two jobs: background help that runs alongside other work, and heavier reasoning that can run separate from cloud infrastructure.
Looking at what this means in practice, the definition of local has moved. An 800-billion-parameter model running outside a data center was implausible not long ago. Unified memory and pooling of multiple machines over high-bandwidth links have shifted that boundary, though the hardware cost stays high. Capacity comes from two linked Studios, but operators must still weigh throughput, latency under concurrent requests, power use and the work of tying such a large file to specific machines.
In my view, the open-weight choice matters as much as the parameter count. Downloadable weights allow inspection, testing on private data, custom setups for agentic evaluation and use with local tools without sending prompts off the device. That suits regulated settings and research groups that need reproducibility. It also leaves guardrails, monitoring and update discipline to the deployer, since a local copy does not get cloud controls automatically.
Worth flagging for inference engineers is the Apple Silicon focus. Tuning for M5 Ultra and unified memory points to work on memory bandwidth, placement and the Metal execution path, rather than a generic checkpoint that runs the same everywhere. How it ports to other stacks and how sensitive it is to precision formats will decide how far it travels beyond the reference setup. Until independent tests repeat Meta's leads on agentic and visual tasks, teams should use those results as a starting point for their own task-specific checks. If those checks hold, more groups gain a way to test heavy agents without leaving their own hardware.
