Meta's Big Release, Through the Rubric
On October 15, Meta put Muse Spark Open on Hugging Face: an 816-billion-parameter mixture-of-experts model, about 42 billion of those active per token, a million tokens of context, and it takes in text, images, audio, and video. The BF16 checkpoint alone is roughly 1.5 terabytes. Whatever else you want to say about Meta, releasing something like this is an enormous act of effort — years of research, a compute bill most countries couldn't cover, and the engineering work of packaging a model this size so that other people can actually run it. That deserves credit before anything else gets said, and I want to say it first, because the rest of this post is just measurement, not judgment.
The rubric from the book asks five separate questions: Access, Permission, Deployment, Transparency, and Reproducibility. Let me walk Muse Spark Open through them one at a time.
Access: yes. The weights are downloadable from Hugging Face, in two formats — a full-precision checkpoint and a 4-bit one built for NVIDIA's newest hardware. You do have to fill out a form with your legal name and organization before the download unlocks, which is a small gate, but it's a gate on paperwork, not on use.
Permission: mostly, with strings. This is where the release stops being "open source" in the strict sense. Meta published it under their own Muse Model License, not Apache 2.0 or MIT. The grant is genuinely generous — you can use the weights commercially, fine-tune them, distill them, build derivative models, and sell what you build. Two conditions shape that generosity. You have to display "Built with Muse Spark Open" in your product. And if you resell access to the model itself as a service and your revenue crosses twenty million dollars in a year, you owe Meta a separate agreement. For almost everyone, that clause never fires. But it's there, and a real open source license doesn't have a revenue trigger in it.
Deployment: yes, emphatically. This is the strongest part of the release. The model card documents how to run it on your own hardware across every major serving stack — vLLM, SGLang, llama.cpp, TensorRT-LLM, even two Mac Studios with MLX. You need serious hardware — the 4-bit version wants around 700 GB of memory — but nothing about the deployment path routes you back through Meta. Download it, run it wherever you like, never send them a token. For me, this is the question that matters most: can I escape the hosted API? With Muse Spark Open, unambiguously yes.
Transparency: partial. The architecture is documented in unusual detail — layer counts, expert structure, the vision and audio encoders, the works. The safety evaluation is published too, with a separate methodology report. But the training data section is a single paragraph: data from public sources, crawled from the internet, acquired from third parties, drawn from Meta's own products, plus synthetic data. No proportions, no source list, no filtering specifics. You know what kinds of things went in; you don't know the recipe.
Reproducibility: no. There's no training code, no data documentation at any real depth, and no path a skilled group could follow to rebuild a substantially equivalent system. Nobody outside Meta can reproduce this model, and honestly, nobody outside Meta could afford to even with the recipe in hand. Training runs at this scale cost more than most companies are worth. I don't think that's a failure of character — it's a fact about the economics of frontier models, and it's why I don't hold missing reproducibility against anyone. But it is the difference between an open release and an open research system.
Put together, Muse Spark Open lands in the middle of the spectrum: open weights with a meaningful custom license and a genuinely open deployment path, partial transparency, limited reproducibility. Strong where it counts for users, quiet where it counts for researchers.
So, but is it open? It's open the way a well-stocked public library with an admission form is open — you can walk in, take the book home, and read it anywhere you want, forever, as long as you keep the author's name on the cover. That's real openness, and Meta spent real money to make it real. It just isn't open source, and the license file is where that difference lives.
Adapted from the draft manuscript of But Is It Open?
