The flagship weights, behind a doorman. Muse Spark Open (October 15) is Meta’s open-weights answer to its own API-only Muse Wonder: 816 billion total parameters, about 42 billion active per token, a million-token context, native multimodal input, quantizations published for NVIDIA, AMD, Intel and Apple silicon on release day. The download asks for more than a click: Meta’s own Hugging Face repositories gate the weights behind a form that takes your full legal name, date of birth and organization identifiers, processed under the Meta privacy policy. It is the first Meta repository to collect identity; every Llama gate before it asked for a checkbox.
The gate is per repository, and the mirrors answer it. The same weights sit ungated at Unsloth’s GGUF rebuild, MLX community conversions and AMD’s official Quark checkpoints. Downloading from Meta puts your identity on file with the ads company; downloading from a mirror skips the form entirely, which is the practical limit of any per-repo permission system.
Fit math. The NVFP4 checkpoint is 484 GB: one 512 GB Mac Studio flagship holds the native build. Unsloth’s ladder starts at 209 GB (IQ1_M, two pooled 128 GB Spark nodes), 311 to 326 GB for the 2-bit class (three-Spark cluster or a 384 to 512 GB pooled Mac). Q4 at 554 GB needs two flagship boxes pooling. With 42B active per token, serving speed tracks a 42B-class model, not an 816B one, on every one of these machines.
The license is Llama tradition carried upmarket. Free use, modification, distillation and distribution; a “Built with Muse Spark Open” attribution requirement on anything built with it; and a separate written agreement once a business’s aggregate AI-model-service revenue with affiliates passes 20 million dollars in any twelve month window. Internal use, research and ordinary products stay free.