Meta Releases Muse Spark Open: 816B Open-Weight Model You Can Run on Two Mac Studios

Open-weight AI just got a lot bigger. On October 15, 2026, Meta released the weights of Muse Spark Open, which it calls “the most capable US open-weight model to date”. It is a natively multimodal model built for coding and agentic work — and Meta says it can run on a desk, not just in a data centre.

The numbers

• ~816 billion parameters in total, ~42 billion active per token
• Mixture-of-experts: 256 experts per MoE layer, 8 selected for each token, 62 layers (Hugging Face)
• Inputs: text, images, audio and video; output: text
• Context window of up to 1 million tokens
• Checkpoints in BF16 (about 1.5 TB) and NVFP4, a 4-bit format for NVIDIA Blackwell GPUs

Meta trained it for long-horizon software work: reading a codebase, writing and debugging code, calling tools. Because it also takes images, audio and video, the same model can read a screenshot, listen to a voice note or watch a screen recording while it works.

“At home” — with a big asterisk

Meta’s headline is that Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. Each machine can give its GPU up to 512 GB of unified memory, so the pair behaves like one computer with a terabyte of memory.

Hugging Face’s own tests show what that means in practice. On a Mac M3 Ultra with 512 GB, a 2-bit llama.cpp quant ran at about 18 tokens/s using 304 GB, and a 3-bit MLX quant reached about 21 tokens/s using roughly 344 GB. Even the 3-bit version needs more than 256 GB of RAM. In the cloud, the BF16 checkpoint runs across four nodes of eight H200 GPUs; the NVFP4 version fits on a single B200 node.

Where to get it

Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs are live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It runs on SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. The licence is the Muse Model License, which Meta describes as having limited commercial terms that let many developers use it for free. One quirk: reasoning can’t be switched off, only set from “low” to “xhigh”.

According to Reuters, it follows Muse Glimmer, a smaller open-weight model Meta released in August.

Rocks take

For a Raspberry Pi running Home Assistant, this is not your next local voice assistant: you need hundreds of gigabytes of memory just to load it. But it matters for local AI in general. A frontier-class model whose weights you can download, quantise and host on hardware you own — or rent by the hour — is a real alternative to closed APIs. The capability claims are Meta’s own; we’ll wait for independent benchmarks. For the smart home, the smaller Muse Glimmer remains the more realistic option.