Muse Spark Open overview

Muse Spark Open is the open-weight model Meta released on October 15, 2026, and the first flagship in the Muse Spark family with downloadable weights. This page collects the published specs. For setup see Run locally; for commercial terms see License explained.

Key specs

Item	Value
Total parameters	816B (Hugging Face shows ~822B including the vision encoder)
Active per token	42B
Architecture	Mixture-of-Experts Transformer, 256 experts, 8 active per token
Layers / hidden size	62 / 8192
Attention	GQA with 128 query heads, 16 KV heads, head dim 64, gated attention, QK-norm
Vocabulary	202,048
Context length	1M tokens
Inputs	Text, image, video, audio (built-in 1.8B Perception Encoder)
Output	Text
Official weight formats	BF16, NVFP4

Why an 816B model can run on a desktop

Because it’s an MoE, each generated token reads only about 42B parameters, so decode speed depends mostly on memory bandwidth. All 816B parameters still have to sit in memory, though, so capacity is the real barrier:

• NVFP4 weights: Meta recommends at least 700GB of VRAM
• BF16 weights: about 2TB
• Meta and Apple’s reference setup: two 512GB M5 Ultra Mac Studios running MLX distributed inference

Try your own combination in the hardware calculator.

Recommended inference settings

• Sampling: temperature=1.0, top_p=1.0, top_k disabled
• Reasoning: reasoning_effort of low, medium, high or xhigh (default xhigh)
• Supported engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, Unsloth

Where to download

• Hugging Face: meta-models/Muse-Spark-Open-816B-A42B (sign in and accept the license)
• Clouds: Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models

FAQ

Q.Is Muse Spark Open open source?

A.It is open-weight. The weights are downloadable from Hugging Face, but they ship under Meta's own Muse Model License rather than Apache or MIT, and commercial use comes with conditions.

Q.Does Muse Spark Open accept images and audio?

A.Yes. The model card lists text, image, video and audio as inputs, with text as the output.