Meta releases Muse Spark Open, an 816B-parameter open-weight model that runs on two Mac Studios

Meta has released the downloadable weights for Muse Spark Open, a natively multimodal model with 816 billion total parameters that the company calls the most capable open-weight model to come out of the United States. The announcement, whose blog post reached the Hacker News front page on 15 October 2026, carries an unusual hardware claim alongside it: this is the first model of its class that can run on a desktop, using two Apple M5 Ultra Mac Studios.

What Meta is shipping

According to the blog post, Muse Spark Open has 816 billion total parameters with 42 billion active, a 1M-token context window and native multimodality. It is aimed at coding and agentic workloads, with Meta highlighting long-running tasks, complex software engineering and multi-step reasoning as its strengths. The company points to a separate evaluation methodology report for details on how those claims were tested.

A frontier-class model on a desk

The headline technical trick is local execution. Using Apple's MLX framework, Meta tuned the model for Mac Studio machines built on the M5 Ultra chip. Because Apple's unified memory architecture lets the integrated GPU address up to 512 GB per machine, two Mac Studios can hold the full model between them, and each chip's 1.2 TB/s of memory bandwidth is fast enough to stream the small slice of the network needed for each token. Meta's framing is that the pair behaves like a single system with a terabyte of unified memory.

To demonstrate the setup, Meta ran agentic workflows entirely on-device using Muse Code, its coding tool powered by the model. The examples included shipping genuine changes to Ice Cubes, an open-source Mastodon client for iOS — bug fixes, new features and a complete build — and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to pick out the strongest moments and crop them vertically.

Availability and tooling

Hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights themselves can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.

Meta lists support across SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA silicon. The release also ships with custom, auto-tunable Helion GPU kernels in SGLang, which Meta says deliver 1.03× to 3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD's MI350. Further customization is possible in PyTorch, with TorchTitan available for tuning the model.

Safety and licensing

Meta acknowledges that open weights carry their own risks, and says Muse Spark Open received additional safety post-training as a result. The model was evaluated under the company's Superintelligence Scaling Framework, which covers the most severe risk categories, and met the thresholds required for release. A model card is available now, with a full evaluation report promised soon.

The weights ship under the Muse Model License, which carries limited commercial terms. Meta says many developers can use the model for free, but the terms are not a fully permissive license in the style of Apache or MIT.

Why it matters

Two things make this release notable. First, it pushes the open-weight frontier: an 816B-parameter multimodal model with a million-token context window is a serious piece of capability to hand out as downloadable files, and the most-capable claim — Meta's own — will invite rapid independent benchmarking. Second, the desktop story matters beyond the demo: if a two-Mac-Studio setup can genuinely serve a model of this size, it narrows the gap between cloud-hosted frontier models and what individuals can run at home, with implications for privacy, cost and vendor lock-in. The restrictive license and the fact that every capability claim so far comes from Meta itself are the caveats to watch as third-party testing arrives.