Muse Spark 1.3 features improved performance in agentic and coding tasks. Deployment has begun today via Muse Code and Meta Model API. Regarding the previously available inference mode, the maximum inference mode is expected to become available as soon as safety testing is completed.
The model is designed to work collaboratively with users to process multiple workflows within a single long thread, allowing for the maintenance of long-term tasks. When given open-ended objectives, it uses tools to generate its own context. It also spontaneously corrects planning deficiencies, records what it has learned, and generates the final deliverables.
Collaboration with users has also been enhanced. If a prompt is ambiguous, the model asks clarifying questions, and if it reaches a standstill, it seeks assistance from the user. It also includes a mechanism to request confirmation before executing critical actions.
Furthermore, multi-tasking capabilities have been improved. Even within complex contexts in a single thread, the model can accurately map input prompts to the correct tasks in response to user instructions or interruptions. The model has been trained to more accurately recognize its own capabilities and limitations, allowing it to determine what it knows and what it does not know instead of hallucinating.