Meta Muse Spark Open: 816B Open-Weight Model
Meta Superintelligence Labs released the open weights of Muse Spark on 15 October 2026: an 816B-parameter mixture-of-experts model with about 42B active per token, a 1M-token context window and native text, image, video and audio input. Meta calls it the most capable US open-weight model to date and says it was optimised with Apple for MLX on M5 Ultra Mac Studios
Tool Overview
Details, capabilities and usage guidance
Features
- ✓ An 816B-total / 42B-active mixture of experts: Meta says Muse Spark Open has roughly 816 billion total parameters and activates about 42 billion per token. It is the full-size model of the Muse series, following Muse Glimmer, the small dense model Meta released in August, and the first time the series takes a mixture-of-experts architecture to this scale
- ✓ A 1M-token context and native multimodal input: Meta says the context window goes up to one million tokens and the model natively brings text, images, audio and video into a single model that generates text, giving developers a common interface for asking questions about very different kinds of input over long documents, screenshots and recordings
- ✓ Trained for long-horizon agentic work: Meta says it is built for coding and agentic work, excelling at long-horizon tasks, complex software engineering and multi-step reasoning. Amazon SageMaker JumpStart also lists the model, which shows it being used both for hosted inference and for self-deployment
- ✓ The Muse Model License: Meta says the weights ship under the Muse Model License, which requires a Built with Muse Spark Open notice in any product that uses them, and a separate written agreement for model-as-a-service providers above $20 million in revenue
- ✓ Optimised for Apple M5 Ultra Mac Studios: Meta says it worked with Apple so Muse Spark Open runs on the MLX framework, with two M5 Ultra Mac Studios paired acting like a desktop supercomputer. The weights are published on Hugging Face with several inference providers offering hosted versions; the BF16 weights are about 1.6TB and the NVFP4 build about 530GB
How to Use
- Find the Muse Spark Open repository under the meta-models organisation on Hugging Face and read the model card first to check the licence and the recommended inference settings (Meta suggests the high or xhigh reasoning effort for complex problem solving, coding and agentic tasks)
- Read Meta's official developer blog post Build with Muse Spark Open for the 816B/42B architecture notes, the 1M-token context and how multimodal input works
- Choose a deployment path for your hardware: self-hosting means dealing with roughly 1.6TB of BF16 weights or the ~530GB NVFP4 build, with MLX on Apple M5 Ultra Mac Studios, while the easier route is a hosted inference provider that already serves the model
- Add the Built with Muse Spark Open notice required by the Muse Model License, and if you resell the model as a service with more than $20 million in revenue, sign a separate agreement with Meta before shipping
FAQ
What is Muse Spark Open?
Meta says Muse Spark Open is the open weights version of Muse Spark, the frontier model behind Meta's hosted Muse products. The weights are an 816-billion-parameter mixture-of-experts transformer with about 42 billion parameters active per token, a one-million-token context, and text, image, video and audio input with text output.
How does it compare with other open-weight models?
Meta says it is the most capable US open-weight model to date. Treat rankings and benchmark scores as claims and check the official model card and independent evaluations; the verifiable facts Meta states are the 816B total / 42B active parameters and the 1M-token context window.
What are the licensing conditions?
Meta says the weights ship under the Muse Model License: any product that uses them needs a Built with Muse Spark Open notice, and model-as-a-service providers above $20 million in revenue need a separate written agreement.
Can an ordinary computer run it?
Meta focuses on Apple M5 Ultra Mac Studios, saying it worked with Apple on MLX so two paired machines act like a desktop supercomputer. The BF16 weights are about 1.6TB and the NVFP4 build about 530GB, which shows self-hosting still needs a large-memory professional machine rather than an ordinary laptop.
Where do I download it?
Meta says the weights are published under the meta-models organisation on Hugging Face and can be downloaded from there, or you can pick an inference provider that already hosts the model; Meta's official Build with Muse Spark Open post is the first document to read.
What is it good for?
Meta says it is built for coding and agentic work, excelling at long-horizon tasks, complex software engineering and multi-step reasoning, and it handles text, image, video and audio input through one interface, which suits things like customer service agents that look things up repeatedly before acting.