Meta Unveils Muse Spark Open, First Public AI Model from Proprietary Series

Introduction

Meta has released Muse Spark Open, the first open-weight AI model from its proprietary series, allowing developers to test its capabilities on complex tasks. The model, which can run on local machines and supports multimodal inputs, is available for testing on platforms such as Arena.ai's Agent Arena and Code Arena.

What is Muse Spark Open?

Muse Spark Open is a multimodal model with 816 billion total parameters. Of those, 42 billion are active. The model also comes with a 1M-token context window. Muse Spark Open can take in images, video, and audio as inputs.

Availability

Muse Spark Open became available on October 15, 2026, through a range of API platforms and cloud marketplaces. API access runs through providers including Baseten and Fireworks AI, among many others. Teams that want the weights themselves can download them from Amazon SageMaker and Google Cloud Model Garden.

Testing and Evaluation

Meta says the model can run on your own infrastructure, with specific optimization for Apple silicon through the MLX framework. The reference setup is two M5 Ultra Mac Studios configured together as a supercomputer with 1TB of unified memory.

Conclusion

Meta Superintelligence Labs claims Muse Spark Open matches frontier closed models on agentic coding benchmarks. Agent Arena and Code Arena put models to work on practical tasks involving coding, web search, and multimedia processing, with live scoring showing how they hold up.