Meta released Muse Spark 1.3, an update to its proprietary AI model, with improved performance on coding and agentic tasks. The new version rolls out today in Muse Code and the Meta Model API, the company announced.
Muse Spark 1.3 is designed for longer-horizon work and multitasking. According to Meta, it asks clarifying questions when prompts are ambiguous, confirms before taking consequential actions, and follows complex, long-form instructions more reliably.
Read more: Meta Rolls Out Primary Device Support for WhatsApp on iPad
The model also maps prompts to the correct tasks in messy contexts, has a better awareness of its own capabilities and limitations, and adapts to user preferences for updates or silent work. Meta says it was trained across diverse harnesses for agentic environments.
A max reasoning mode is planned but will arrive only after safety testing, with no specific date given. Pricing details for the API were not disclosed.
Muse Spark 1.3 follows the original Muse Spark, Meta's first major proprietary AI model, which marked a strategic reversal from its open-source Llama line. That model arrived after Meta poured $14 billion into talent and infrastructure at Scale AI, and at the time Meta said it would run its digital assistant across its apps and Ray-Ban glasses.
The earlier model was developed by Meta Superintelligence Labs, led by Alexandr Wang, and took nine months to build using distillation from third-party models. At launch, Meta said it was not as capable as ChatGPT or Gemini in some areas, and shares jumped 6% after the announcement.
