Meta Adds Max Reasoning to Muse Spark 1.3

Meta's September 4 post made the strongest reasoning setting of its Muse Spark 1.3 model, called max, available to everyone on Muse Code, Meta's coding agent, and the Meta Model API. A reasoning setting controls how long the model works through a problem before it answers, and higher settings use more tokens per task. When Meta released Muse Spark 1.3 two days earlier, the setting the public could use was xhigh, while max was in limited preview for Meta's partners, according to Artificial Analysis, an independent benchmarking firm.

Artificial Analysis tested that preview and found gains over xhigh. On its Intelligence Index, max scored 62 against 61 for xhigh, and on its Coding Agent Index, run inside Muse Code, max scored 68 against 64, which the firm ranked second, behind only Claude Opus 5 running in Claude Code, which also scored 68. The firm attributed max's better agentic results to more turns and more reasoning, for example 62 percent more reasoning on its GDPval-AA v2 test than xhigh, which means each task takes longer and uses more tokens. Alexandr Wang, Meta's chief AI officer, said in the post Meta quoted that the company sees significantly stronger coding and agentic performance with max and recommended it even to people who had tried high or xhigh. Meta's post gives no price for max, which Artificial Analysis said was not yet public, and the firm's figures were measured on the partner preview before the public release.