Muse Spark Open arrives on Together’s hosted service

Request a response through the API

Together’s announcement shows a Python example that sends a user’s request to the model through its API, a programmatic interface. The example asks for a code refactor, then prints the returned message. It shows how to obtain a response, not how to execute the proposed code. A browser playground offers another way to try the model.

Downloadable weights, too

Meta’s release announcement also makes the model weights available for download: developers are not limited to hosted access. Meta lists several supported inference engines, the software used to run the model, including vLLM and SGLang. It also describes further customization using PyTorch.

For local use, Meta describes two M5 Ultra Mac Studios running through MLX. Its reported on-device demonstrations include Muse Code making bug fixes and feature changes to the Ice Cubes iOS app and completing a build. These are demonstrations reported by Meta, not independent tests here.

Local deployment has requirements

The model card lists text, images, video and audio as inputs, but text as output. For GPU deployment on your own system, it specifies at least 700 GB of video memory for NVFP4 and 2 TB for BF16, alongside a supported inference framework.

Meta warns that the model can still give inaccurate answers and make mistakes in multi-step reasoning. For agents taking real-world actions, it recommends extra safeguards, such as human confirmation before irreversible actions. This is a deployment recommendation, not a guarantee of correct results.