Meta has released the weights for Muse Spark Open, an 816B total parameter (42B active) mixture-of-experts model with a 1M-token context window, billing it as the most capable US open-weight model available. The model targets coding, agentic workflows, and multi-step reasoning tasks, and is available immediately through seven API platforms and four model repositories including Hugging Face. The headline consumer pitch is desktop inference: two Apple M5 Ultra Mac Studios, each with 512GB of unified memory, can hold the full model and run it at usable speeds thanks to 1.2 TB/s memory bandwidth per chip. Meta demonstrated real agentic workflows on-device — shipping iOS changes to an open-source Mastodon client and driving DaVinci Resolve end-to-end. This is genuinely new territory for local model running at this scale. The infrastructure story is broader. Muse Spark Open runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth, with optimizations spanning AMD, Apple, Intel, and NVIDIA silicon. Custom Helion GPU kernels in SGLang deliver 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, with a 3% end-to-end speedup on AMD MI350. The model is tunable via PyTorch's TorchTitan. But 'open' deserves scrutiny. The release uses the Muse Model License with 'limited commercial terms,' not Apache 2.0 or MIT. Meta controls what 'many developers' means in practice — the license permits free use for some but not all commercial applications. This is the same open-weight-not-open-source pattern Meta established with Llama: weights are downloadable, but the license is proprietary and Meta retains control over the terms of use. The Apple coupling is strategic on both sides. The M5 Ultra Mac Studio starts at roughly $8,000, and you need two of them. Meta gets a flagship demonstration of local inference that most developers will experience through API partners instead. Apple gets a reason to sell its most expensive desktop silicon to AI developers. The 'run it at home' narrative is real but economically gated — this is not democratization at the $500 GPU level. Safety post-training and evaluation against Meta's Superintelligence Scaling Framework are mentioned but details are deferred to a forthcoming evaluation report. The model card exists; the full safety evaluation does not yet. For a model released under the banner of 'personal superintelligence for everyone,' the safety documentation trails the marketing. The competitive dynamic is clear: Meta is using open weights as a distribution strategy to ensure its model ecosystem becomes the default layer beneath thousands of applications, while cloud inference partners do the heavy lifting of serving most actual queries. The generative value is real — developers get a powerful model they can fine-tune and deploy. The extraction is in the ecosystem lock-in: every application built on Muse Spark Open deepens Meta's position as the de facto open-weight standard, which is a moat dressed as a gift.
