Meta released the weights for Muse Spark Open on October 15, calling it the most capable US open-weight model to date. The headline spec is a natively multimodal mixture-of-experts with 816B total parameters, 42B active per token, and a 1M-token context window, built for coding and long-horizon agentic work (Meta AI Research blog, Oct 15). The detail that changes my planning is not the parameter count. It is the hardware path: Meta says it is the first model in this class to run on your desktop, using two Apple M5 Ultra Mac Studios and the MLX framework.
The memory math is the real story
The way it fits is Apple's unified memory, not a GPU cluster. Each M5 Ultra Mac Studio can expose up to 512 GB to its integrated GPU, so two machines hold the entire model, and the source frames them as behaving like one machine with about a terabyte of unified memory. The M5 Ultra's 1.2 TB/s of memory bandwidth is what makes it viable: the model reads only a small slice of the 42B active parameters per token, so the two boxes together keep up where the slice is the bottleneck.
For comparison, my local stack has been single-machine for a long time. As covered in the LFM2.5-24B-A2B on-device post, the constraint at that scale is one device. This is a different regime: the model only runs on a two-machine pair, and the 42B-active / 816B-total split is what lets a MoE ride on 1 TB of unified memory at all.
What actually runs, and what it's tuned for
The source names the inference engines it runs on: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. For Apple specifically, it pulls weights straight from Hugging Face through an MLX adapter.
The GPU side ships with custom, auto-tunable Helion kernels in SGLang, which Meta reports as 1.03-3.95x kernel-level speedups on NVIDIA and AMD, plus a 3% end-to-end speedup on an AMD MI350. Those kernel multipliers are GPU numbers; the Mac Studio path runs on unified-memory MLX, so the two performance stories are separate and I would not read the 3.95x as a figure to expect on Apple silicon.
On-device, Meta tested agentic workflows through Muse Code running entirely local, including shipping real changes to Ice Cubes (an open-source Mastodon client) — bug fixes, new features, full build — and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find and vertically crop the strongest moments.
Where you get it, and the catch
| Weights (download) | Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, Microsoft Foundry |
| Hosted inference | Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together AI |
| Fine-tune | PyTorch via TorchTitan |
| License | Muse Model License, limited commercial terms |
The weight download is on Hugging Face at the meta-models/muse-spark-open collection. If you want to customize it, the documented route is PyTorch with TorchTitan.
The catch is the entry price in hardware: two M5 Ultra Mac Studios is a real capex decision, not a hobbyist box, and the "free" framing is the model license, not the compute. I'd also watch the evaluation report, which the source says is available soon. Safety was evaluated under the Meta Superintelligence Scaling Framework and the model has extra post-training for open-weight release, but I would read the full report before treating the benchmark numbers as settled. As a reference point, my Mistral Large 4 open-weights post covered a similar open-weight frontier tier at 1T parameters — Muse Spark Open lands in that same conversation but is explicitly aimed at desktop, on-device serving.
Sources: