Muse Spark 1.3 (xhigh)

Muse Spark 1.3 is a multimodal reasoning model developed by Meta Superintelligence Labs (MSL), released in September 2026. It is a frontier-level system optimized for long-horizon agentic workflows, advanced coding tasks, and autonomous computer use. The model serves as the backbone for Muse Code, Meta's terminal-based coding agent, and is available through the Meta Model API.

Technically, Muse Spark 1.3 introduces multiple reasoning tiers, including xhigh for general availability and max for high-complexity reasoning tasks. The model is designed to be more collaborative than previous versions, frequently asking clarifying questions and confirming consequential actions before execution. It demonstrates improved efficiency over its predecessor, utilizing approximately 25% fewer tokens and 20% fewer tool calls for equivalent tasks.

Capabilities and Performance

The model features a native 1,048,576 token context window, enabling it to manage extensive codebases and sustain state across long-running multi-agent threads. On industry benchmarks, Muse Spark 1.3 has shown competitive performance, scoring 75.4% on the DeepSWE 1.1 software engineering test and maintaining high accuracy (98.5%) on long-context retrieval (MRCR) benchmarks. It is specifically tuned for tool-use scenarios, matching or exceeding contemporary models like Claude Opus 5 in terminal-based coding benchmarks.

Beyond text-based tasks, Muse Spark 1.3 supports multimodal perception, allowing it to interpret visual data such as complex diagrams and user interface patterns. It has also been tested for real-world infrastructure provisioning, successfully performing tasks like deploying AWS CloudFront and S3 resources from single prompts. While initially launched as a proprietary model, Meta has confirmed plans for a future open-weight release of the Muse Spark family.