Meta’s newest AI model is now in front of the toughest audience in tech: other developers.
Muse Spark Open, the first open-weight release from Meta’s Muse Spark series, is now available in Arena.ai’s Agent Arena and Code Arena. There, users can test it on complex tasks and score how it actually performs. Meta announced the model on October 15, 2026. Arena.ai confirmed the integration and live testing on launch day.
What Meta actually built
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.
Meta is pitching the model at three jobs in particular: long-horizon agentic workflows, complex coding, and multi-step reasoning. On the coding side, Meta says it supports workflows such as analyzing entire repositories and handling complicated refactors that span multiple files.
Where you can get it
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.
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.
Meta also says the model went through safety post-training, which according to Meta lets Muse Spark Open operate on a standalone basis without outside guardrails bolted on.
Why Arena is the real test
Muse Spark Open builds on Meta’s earlier Muse Spark models, which were proprietary. This release marks the series’ first move into open weights.
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.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.