Meta has released the weights for Muse Spark Open, which it calls "the most capable US open-weight model to date," in a blog post dated Oct. 15, 2026.

The specs

  • 816 billion total parameters, 42 billion active per token (mixture of experts).
  • 1 million-token context window.
  • Natively multimodal: takes text, images, audio and video, and outputs text, Hugging Face's launch post says.
  • 256 experts per MoE layer, 8 selected per token, 62 layers, per Hugging Face.

On your desk, sort of

Meta says it is the first model in its class to run on a desktop setup: two Apple M5 Ultra Mac Studios with up to 512 GB of unified memory each, using Apple's MLX framework.

Meta says it tested on-device agent workflows, including shipping real iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end.

Where to get it

API access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry, Meta said.

Hugging Face says Meta trained it for agentic coding: reading a codebase, writing and debugging code, and calling tools. It follows Muse Glimmer, a small dense model Meta released in August.

The "most capable" claim is Meta's own. For other recent open releases see DeepSeek V4.1-Flash-1017 and Nvidia Nemotron 3.5.