Meta released Muse Spark Open on [date], saying it is trained for agentic coding workflows with a focus on first-attempt accuracy and understanding of large codebases. It is the company's first major model update since the release of its earlier coding-focused models.
Meta reported 816 billion total parameters, measured in a mixture-of-experts (MoE) architecture, with 42 billion active for each token. That compares with the smaller parameter counts of its previous models.
Muse Spark Open is built on a reasoning model architecture and targets coding, debugging, and long-running tasks. Availability begins now, initially for developers using Amazon SageMaker JumpStart.
"Muse Spark Open is designed to understand and work with large codebases, enabling developers to refactor, debug, and generate code with the full context of a project," said Meta’s representative. The model accepts text, images, documents, audio, and audio-video as input and returns text.
The announcement follows Meta’s broader push into open-source AI models. Meta framed the release as a step toward more accessible and versatile AI tools for developers.
Meta did not say how the model will perform on tasks beyond coding, and it raised the question of how well it will handle multimodal reasoning in real-world scenarios. The model is now available for developers to test and deploy.
Source: awsml
