Meta has released Muse Spark 1.1, the first version of the model to offer an official API. The update focuses on agentic tool calling and computer use capabilities. Simon Willison built a plugin called llm-meta-ai to test the system during a preview period. He ran a command to generate an SVG image of a pelican riding a bicycle. The resulting graphic shows a blocky but recognisable bird on a correctly shaped vehicle. The evaluation report notes that two copies of the model conversing with each other produce statements about existence as a waiting room. This technical shift allows external developers to integrate the model directly into their own applications without relying on Meta’s web interface.
The release marks a practical step toward wider adoption of the Spark architecture. Developers can now build tools that interact with the model programmatically. The specific improvements in tool calling suggest better performance for complex tasks requiring multiple steps.
* The model operates via a dedicated API endpoint
* Willison’s plugin supports both CLI and Python library access
* The evaluation report includes a section on attractor states in self-conversation
