Meta Muse Spark 1.1 Performance Overview

The Breakthrough

Meta has launched Muse Spark 1.1, a multimodal reasoning model via the Meta Model API that demonstrates high proficiency in agentic task orchestration and complex mathematical problem-solving while maintaining a 1-million-token context window.

Performance and Benchmarking

The model achieved a score of 48 out of 70 (68.57%) on the author's KingBench, which evaluates UI generation, 3D graphics, math, and agentic capabilities. While the model successfully solved a complex permutation counting problem that stumped several other frontier models, it showed significant inconsistency in coding tasks, particularly regarding logic implementation and file-system awareness.

Agentic and UI Capabilities

• Agentic Workflow: The model successfully generated a dataset, performed local fine-tuning on a Gemma 2B model, and deployed a functional web UI that served facts from the fine-tuned model.
• UI Design: The model consistently produced high-quality visual designs, even when the underlying functional logic or animation state management failed to meet requirements.
• Tool Use: The model performed well on structured tool-based workflows, though it struggled with parallel agent setups where it frequently ignored existing files and overwrote data from other sessions.

Context

Meta entered the frontier model space with Muse Spark 1.1, positioning it as a tool for reasoning, coding, and computer use. Priced at $1.25 per million input tokens and $4.25 per million output tokens, it is marketed as a cost-effective alternative for agentic tasks. Despite strong performance on specific benchmarks like MCP Atlas, the model currently lacks the reliability required for complex, multi-file coding projects compared to established alternatives like Opus 4.8 or Fable 5.