Meta's Muse Spark Open hits 5M downloads - and Apple is bringing it to Mac silicon

The release. Meta has put its open-weight frontier model on the map: Muse-Spark-Open-816B-A42B is a multimodal Mixture-of-Experts with roughly 800B total parameters and 42B active per token. Meta's Hugging Face blog pitches it as something you can "own, customize, and scale"-language that reads less like an open-source PR and more like a product strategy. The bet: if the weights are yours, the moat is the ecosystem around them.

Adoption is already ahead of the hype. The Muse family has passed 5M downloads, a faster ramp than ChatGPT or Claude at a comparable stage, according to the Hacker News front-page roundup. For an open-weight model family, that is a fast start.

On Mac silicon. The story that traveled today: Apple's ML Research is bringing Muse Spark Open to Mac Studio via MLX-per the ML Research post, a first-tier lab is now shipping open weights tuned for local Apple silicon. It was one of the top AI stories on Hacker News today. The direction of travel is clear: frontier-class architectures that a workstation can actually serve, with a first-tier lab doing the tuning.

Open weights are winning the adoption war

The release is not an island. In the same news cycle, Hugging Face's blog announced Beam 501B, a new 501B open model. Beam 501B landing in the same window as Muse Spark Open is a statement about where frontier labs' attention is going, and the pattern behind it is the one today's digests keep repeating: open-weight adoption is quietly outpacing the closed giants.

The closed side is answering with price. Last Week in AI's report has OpenAI launching GPT-6 Sol/Luna at lower cost while Anthropic shipped Claude Opus 5.5 with stricter cyber safeguards-a price/performance race that is compressing margins for everyone below the frontier. Open weights plus falling API prices is a pincer on the middle of the market.

What it means if you run things locally

Honest math first: the flagship will not fit in 128GB of memory even quantized. The local story is the derivative-a distilled small sibling with the same 42B-active behavior is the shape to watch, and it is the only version of this model family that will ever touch a desktop. The trending charts' read is consistent: 800B-class open weights are aspirational until a small cut ships.

Until then, today's genuinely local picks from the Hugging Face charts are MiniCPM-V-4.7-35B-A3B-a 35B vision-language MoE with just 3B active parameters, fast enough for CPU and 32GB-class hardware-and LTX-2.5, the image-to-video diffusion model that hit ~1.9M downloads this week and is the most local-friendly video model on the market.

Signal to watch: whether Meta or Apple ships a distilled Muse Spark cut within a few weeks-that, not the 816B flagship, is what decides whether open-weight frontier models move from datacenter to desktop.