Meta Releases Muse Spark 1.3

Meta Superintelligence Labs has released Muse Spark 1.3, enhancing agent performance and efficiency. This update dramatically improves information retrieval accuracy within a 1M token context from 55.5 points to 98.1 points, marking a breakthrough in long-context processing capabilities.

Performance and Efficiency Improvements

• Coding Efficiency: Reduced tool call counts by approximately 20% and token usage by approximately 25% for the same tasks, lowering operational costs.
• Agent Capabilities: The Agentic IF Index score rose from 46.2 points to 57.8 points, with improved safety and contextual understanding, such as asking clarifying questions for ambiguous instructions or confirming before irreversible actions.
• Benchmarks: Scored 75.4 points on DeepSWE v1.1, slightly surpassing Claude Opus 5, but trailed Opus 5 on some metrics like GDPVal-AA v2.

Pricing and Accessibility

• API Pricing: Input is priced at $0.10 and output at $0.20 per 1M tokens (based on data contributors), maintaining competitive pricing.
• Open Weights: Currently, only the distilled 30B parameter model Muse Glimmer (Apache 2.0) has been released; the release schedule for open weights of the main model is undetermined.