Meta’s 816B Muse model reaches desktops — with a hundreds-of-gigabytes catch

Meta released Muse Spark Open on October 15 as an 816‑billion‑parameter open‑weight model. Its mixture‑of‑experts design activates about 42 billion parameters for each token, so every response uses only part of the full network even though the complete weights still have to be stored. The model accepts text, images, video and audio and supports a context window of up to one million tokens. Meta says it trained the model for long‑running agent tasks such as reading a codebase, planning a change, calling tools and checking the result. “Open weights” means developers can download and run the model files under Meta’s Muse Model License. It does not mean the model is small, free to operate or released under an unrestricted open‑source license.

Meta’s launch guide says the BF16 reference weights occupy about 1.6 terabytes. Its NVFP4 checkpoint is roughly one‑third that size at about 530GB and can run across two Mac Studios with 512GB of unified memory each or four 256GB systems connected over Thunderbolt 5 and RDMA. More aggressive community quantization lowers the floor. The same guide says an approximately 2‑bit GGUF build can fit on a single workstation with about 300GB of combined RAM and VRAM, while its summary table describes a 256GB setup. Those numbers refer to different compressed packages and leave little room for other workloads. The headline consequence is therefore narrower than “runs on a Mac.” This is desktop‑class deployment for unusually high‑memory workstations, not ordinary use on a 16GB or 32GB laptop.

Sparse mixture‑of‑experts models route each token through selected parts of the network. Activating 42B parameters reduces computation relative to using all 816B for every token, but storage and memory still have to accommodate the full checkpoint or a compressed version. On Apple hardware, Meta says MLX tensor parallelism divides every layer across machines so each computer stores and processes a share of the weights. Small intermediate results cross the high‑speed connection while the weights remain local to each machine. That architecture can make local inference possible without a conventional GPU server, but the operational cost moves into large unified‑memory systems, fast interconnects, power and setup expertise. Sparse is a compute property, not a promise of consumer‑device affordability.

Meta lists launch support from hosted providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It also lists the model in cloud catalogs from Amazon, Google and Microsoft. Hosted access lets a team evaluate prompts, tool use and latency before buying hardware. It also transfers some infrastructure work to the provider, while introducing usage charges, data‑handling terms and service availability that need separate review. The right comparison is total workload cost. A local system may suit steady sensitive workloads; an API can be cheaper for sporadic experiments. The model release itself does not prove which option wins for a particular team.

Muse Spark Open ships under the custom [Muse Model License](https://huggingface.co/meta‑models/Muse‑Spark‑Open‑816B‑A42B‑NVFP4). Anyone planning redistribution, commercial deployment or a sensitive application should read the actual license and model card rather than infer permissions from the phrase “open weights.” Meta’s performance and partner claims come from the release materials. Independent testing will be needed to establish how the model behaves across hardware, long contexts, languages and real agent workflows. This article does not compare benchmark scores or treat launch‑day partner quotes as independent validation. The release still marks a meaningful industry shift: a frontier‑scale model can move outside a single vendor’s hosted service. The catch is that control over the weights arrives with hundreds of gigabytes of memory, a specialized runtime and responsibility for operating the system safely. *Image: Meta’s official Muse Spark Open model‑card artwork on Hugging Face, reproduced without generative alteration or artificial upscaling.*