Meta Announces Muse Spark 1.3, Enhancing Coding and Agent Performance

For agent-based applications, the model has improved its ability to sustain long-term tasks. New features include the ability to ask clarifying questions when faced with ambiguous instructions and the capacity to request help from the user when stuck. Meta explained that the model has been trained to recognize the limits of its own capabilities and communicate when it has reached a standstill.

Improvements are also evident in coding performance. According to comparisons by Meta engineers, tool calls decreased by approximately 20% and token consumption decreased by approximately 25% compared to the previous model, "Muse Spark 1.2." This is said to reduce redundancy and suppress unnecessary interactions.

A published score sheet showed comparison results with competing models. In three coding-related categories, Muse Spark 1.3 outperformed OpenAI's "GPT-5.6 Sol" and Anthropic's "Claude Opus 5," recording the highest scores. On the other hand, Claude Opus 5 showed the highest scores in four out of six agent-related categories.

CEO Mark Zuckerberg stated that this is the "biggest leap yet" in coding and agent-like tasks. He noted that Meta plans to release larger models in the future, as well as the open-weight release of "Muse Spark 1.2" with safety fine-tuning.